No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Machine and deep learning for sport-specific movement recognition: a systematic review of model development and performance
Abstract
Objective assessment of an athlete's performance is of importance in elite sports to facilitate detailed analysis. The implementation of automated detection and recognition of sport-specific movements overcomes the limitations associated with manual performance analysis methods. The object of this study was to systematically review the literature on machine and deep learning for sport-specific movement recognition using inertial measurement unit (IMU) and, or computer vision data inputs. A search of multiple databases was undertaken. Included studies must have investigated a sport-specific movement and analysed via machine or deep learning methods for model development. A total of 52 studies met the inclusion and exclusion criteria. Data pre-processing, processing, model development and evaluation methods varied across the studies. Model development for movement recognition were predominantly undertaken using supervised classification approaches. A kernel form of the Support Vector Machine algorithm was used in 53% of IMU and 50% of vision-based studies. Twelve studies used a deep learning method as a form of Convolutional Neural Network algorithm and one study also adopted a Long Short Term Memory architecture in their model. The adaptation of experimental set-up, data pre-processing, and model development methods are best considered in relation to the characteristics of the targeted sports movement(s).
... There are already various reviews about the application of machine and deep learning techniques for human motion analysis. Ref. [7] have already analysed the application of machine learning algorithms in the field of sport-specific movement recognition in a systematic review for the period of January 2000 to May 2018. The authors analysed in-field automatic sport-specific movement detection and recognition with IMU and vision-based data. ...
... The aim of this work is to identify the classification methods used for human motion analysis in sports in the form of a scoping review and to analyse whether and how the time series structure of the data is considered. Compared to [7], we focus exclusively on classification tasks and IMU data, do not only include machine or deep learning algorithms, and also include in-lab studies. We mainly discuss the fifth step of the ARC and investigate the methodological and statistical specifications of human motion data classification in sports across all disciplines. ...
... We adapted the keyword string of [7] by focusing on general (individual) motion classification in sports, but did not concentrate on outcome predictions of matches and games. ...
Inertial measurement units (IMU) are widely used in sports applications to digitise human motion by measuring acceleration and rotational velocity in three-dimensional space. A common machine learning problem is the classification of human motion primitives from IMU data. In order to investigate the classification methods used in the existing literature and to analyse whether and how the time-dependent data structure is considered in the classification process of motion data analysis in sports, a scoping review was conducted. Based on a keyword search, articles from 2010 to 2021 were extracted, and 93 articles were relevant for data extraction. Over- and undersampling of data and data augmentation techniques were rarely used. The classification methods applied can be divided into three main branches: classic machine learning and deep learning models, threshold-based approaches, and dynamic time warping. The most often applied algorithms were support vector machines (SVM), followed by neural networks and k-nearest neighbours. In comparative works, when more than one classifier was applied, random forests, neural networks, boosting models and SVM were found to be the methods that achieved the highest accuracy. If the time-dependent data structure was taken into account, it was incorporated either within the models, for example, by using long-short-term memory models or within the feature calculation step by using rolling windows with an overlap, which was the most common method of considering the time dependency of the IMU data.
... This study seeks to contribute to the existing literature on the validity and reliability of computer vision applications for measuring joint angles, particularly in elite athletes. By establishing the accuracy and consistency of these tools, we can pave the way for their wider use in clinical practice, ultimately improving patient outcomes and enhancing the performance of athletes [15,[19][20][21][22]. ...
... It seems that technological advances could help make the measurement of joint angles faster and more reliable, and there is also a trend that markers are no longer needed [20,22]. By harnessing the capabilities of computer vision applications and machine learning, clinicians could gain valuable insights into the risk factors associated with different activities or the elderly, facilitating the development of targeted injury prevention strategies [21,37]. Additionally, objective measures of changes in joint health over time could be obtained, enabling the evaluation of training and rehabilitation programs. ...
... Additionally, objective measures of changes in joint health over time could be obtained, enabling the evaluation of training and rehabilitation programs. With the growing emphasis on evidence-based practice in healthcare, the use of objective, quantifiable measures provided by computer vision-based applications could enhance the accuracy of clinical decision-making, ultimately leading to improved patient outcomes [19,21,38]. ...
Background:
In handball, the kinematics of the frontal plane seem to be one of the most important factors for the development of lower limb injuries. The knee valgus angle is a fundamental axis for injury prevention and is usually measured with 2D systems such as Kinovea software (Version 0.9.4.). Technological advances such as computer vision have the potential to revolutionize sports medicine. However, the validity and reliability of computer vision must be evaluated before using it in clinical practice. The aim of this study was to analyze the test-retest and inter-rater reliability and the concurrent validity of a beta version app based on computer vision for the measurement of knee valgus angle in elite handball athletes.
Methods:
The knee valgus angle of 42 elite handball athletes was measured. A frontal photo during a single-leg squat was taken, and two examiners measured the angle by the beta application based on computer vision at baseline and at one-week follow-up to calculate the test-retest and inter-rater reliability. A third examiner assessed the knee valgus angle using 2D Kinovea software to calculate the concurrent validity.
Results:
The knee valgus angle in the elite handball athletes was 158.54 ± 5.22°. The test-retest reliability for both examiners was excellent, showing an Intraclass Correlation Coefficient (ICC) of 0.859-0.933. The inter-rater reliability showed a moderate ICC: 0.658 (0.354-0.819). The standard error of the measurement with the app was stated between 1.69° and 3.50°, and the minimum detectable change was stated between 4.68° and 9.70°. The concurrent validity was strong r = 0.931; p < 0.001.
Conclusions:
The computer-based smartphone app showed an excellent test-retest and inter-rater reliability and a strong concurrent validity compared to Kinovea software for the measurement of the knee valgus angle.
... Machine learning software algorithms have demonstrated reasonable accuracy (>80%) in the identification of various postures and movements in adults [14]. There has been particular interest in applying machine learning to identify specific postures and movements in sporting contexts [15] and daily activity monitoring for people with movement impairments [16]. Collectively, this research has shown that adult postures and movements can be identified with reasonable accuracy (>80%) in a range of different contexts [14][15][16]. ...
... There has been particular interest in applying machine learning to identify specific postures and movements in sporting contexts [15] and daily activity monitoring for people with movement impairments [16]. Collectively, this research has shown that adult postures and movements can be identified with reasonable accuracy (>80%) in a range of different contexts [14][15][16]. Whilst several studies have focused on the accurate identification of postures and movements using accelerometry data collected on young children, this information is yet to be synthesised. ...
Given the importance of young children’s postures and movements to health and development, robust objective measures are required to provide high-quality evidence. This study aimed to systematically review the available evidence for objective measurement of young (0–5 years) children’s posture and movement using machine learning and other algorithm methods on accelerometer data. From 1663 papers, a total of 20 papers reporting on 18 studies met the inclusion criteria. Papers were quality-assessed and data extracted and synthesised on sample, postures and movements identified, sensors used, model development, and accuracy. A common limitation of studies was a poor description of their sample data, yet over half scored adequate/good on their overall study design quality assessment. There was great diversity in all aspects examined, with evidence of increasing sophistication in approaches used over time. Model accuracy varied greatly, but for a range of postures and movements, models developed on a reasonable-sized (n > 25) sample were able to achieve an accuracy of >80%. Issues related to model development are discussed and implications for future research outlined. The current evidence suggests the rapidly developing field of machine learning has clear potential to enable the collection of high-quality evidence on the postures and movements of young children.
... Sports video analysis aims to provide objective measures to compare the athlete's performance, e.g. during training. For a given sports performance, video analysis can help to improve the athlete's technique with real-time feedback [9]- [13]. Thus, it assists sport trainers in technique analysis, thereby increasing the overall efficiency of the coaching process [14]. ...
... Given 4 points, a homography could be solved with a unity constraint to cover the last degree of freedom, i.e. (12), (13). Via the COM tracking from section V-B, we can easily identify multiple points in the image plane ((x, y) i ) that are not lying on one line. ...
The tremendous progress of deep convolution neural networks has shown promising results on the classification of various sports activities. However, the accurate localization of a particular sports event or activity in a continuous video stream is still a challenging problem. The accurate detection of sports actions enables the comparison of different performances, objectively. In this work, we propose the DiveNet action localization module to detect the springboard diving sports action in an unconstrained environment. We used Temporal Convolution Network (TCN) over a backbone feature extractor to localize diving actions, with low latency. We estimate the divers center of mass (COM) trajectory and the peak dive height using the temporal demarcations provided by the action localization step via the projectile motion formula. In addition, we train a DiveNet pose regression network, which extends the Unipose architecture with direct physical parameter estimation, i.e COM and 2D joint keypoints. We propose a new homography computation method between the diving motion plane and the image-view for each dive. This enables the representation of physical parameters in metric scale, without any calibration. We release the first publicly available diving sports video dataset, recorded at 60 Hz with a static camera setup for different springboard heights. DiveNet action localization achieves an accuracy of 95% with a single frame latency (< 25 ms). The DiveNet pose regression model shows competitive results around 70% PCK on different diving pose datasets. We achieve COM accuracy of 6 pixels, dive peak height sensitivity of 20 cm and mean joint angle errors around 10 degrees.
... In competitive and professional sports, all athletes are monitored in nearly every game and, if possible, also during training. The monitoring provides data that can be analysed to further improve the performance of individual athletes or the team, but it can also deliver information about opposition teams, their tactics and strategy, strength and weaknesses, etc. [1][2][3][4][5]. As the amount of available data is too large to be processed efficiently by coaches and analysts, the state of the art in the analysis of such data comprises a mixture of computer-aided and human analysis and evaluation [2,[5][6][7][8]. ...
... The monitoring provides data that can be analysed to further improve the performance of individual athletes or the team, but it can also deliver information about opposition teams, their tactics and strategy, strength and weaknesses, etc. [1][2][3][4][5]. As the amount of available data is too large to be processed efficiently by coaches and analysts, the state of the art in the analysis of such data comprises a mixture of computer-aided and human analysis and evaluation [2,[5][6][7][8]. The computer-aided part of the analysis is mostly based on modern algorithms, e.g., methods of machine learning [9][10][11][12][13], though before any analysis can be carried out, the data has to be gathered. ...
In order to train receivers in American football in a targeted and individual manner, the strengths and weaknesses of the athletes must be evaluated precisely. As human resources are limited, it is beneficial to do it in an automated way. Automated passing machines are already given, therefore the motivation is to design a computer-based system that records and automatically evaluates the athlete’s catch attempts. The most fundamental evaluation would be whether the athlete has caught the pass successfully or not. An experiment was carried out to gain data about catch attempts that potentially contain information about the outcome of such. The experiment used a fully automated passing machine which can release passes on command. After a pass was released, an audio and a video sequence of the specific catch attempt was recorded. For this purpose, an audio-visual recording system was developed which was integrated into the passing machine. This system is used to create an audio and video dataset in the amount of 2276 recorded catch attempts. A Convolutional Neural Network (CNN) is used for feature extraction with downstream Long Short-Term Memory (LSTM) to classify the video data. Classification of the audio data is performed using a one-dimensional CNN. With the chosen neural network architecture, an accuracy of 92.19% was achieved in detecting whether a pass had been caught or not. The feasibility for automatic classification of catch attempts during automated catch training is confirmed with this result.
... Measuring sports movement during training and competition allows monitoring athletes' performance and their risk of injury (Camomilla, Bergamini, Fantozzi, & Vannozzi, 2018;Cust, Sweeting, Ball, & Robertson, 2019). Performance monitoring is relevant to assess motor capacity and physical demand, as well as to analyze technique and how technique impacts performance (Camomilla et al., 2018). ...
... Machine learning is a powerful tool where the recorded data are used to make inferences about the data themselves (Bishop, 2006). Supervised learning, unsupervised learning, and reinforcement learning are increasingly investigated in human movement analysis, for example, to classify normal and pathological gait, to map IMU data to biomechanical variables, to discover or cluster movement patterns, or to learn controllers that drive biomechanical models (Cust et al., 2019;Ferber, Osis, Hicks, & Fig. 1. Overview of "in the wild" movement recording and analysis using machine learning. ...
Recent advances in wearable sensing and machine learning have created ample opportunities for “in the wild” movement analysis in sports, since the combination of both enables real-time feedback to be provided to athletes and coaches, as well as long-term monitoring of movements. The potential for real-time feedback is useful for performance enhancement or technique analysis, and can be achieved by training efficient models and implementing them on dedicated hardware. Long-term monitoring of movement can be used for injury prevention, among others. Such applications are often enabled by training a machine learned model from large datasets that have been collected using wearable sensors. Therefore, in this perspective paper, we provide an overview of approaches for studies that aim to analyze sports movement “in the wild” using wearable sensors and machine learning. First, we discuss how a measurement protocol can be set up by answering six questions. Then, we discuss the benefits and pitfalls and provide recommendations for effective training of machine learning models from movement data, focusing on data pre-processing, feature calculation, and model selection and tuning. Finally, we highlight two application domains where “in the wild” data recording was combined with machine learning for injury prevention and technique analysis, respectively.
... Using Dijkstra's shortest path optimization 44 , the ideal path through each keypoint distribution is determined (7). Finally, the refined pose is composed using the optimised keypoints only (8). www.nature.com/scientificreports/ ...
For most applications, 2D keypoint detection works well and offers a simple and fast tool to analyse human movements. However, there remain many situations where even the best state-of-the-art algorithms reach their limits and fail to detect human keypoints correctly. Such situations may occur especially when individual body parts are occluded, twisted, or when the whole person is flipped. Especially when analysing injuries in alpine ski racing, such twisted and rotated body positions occur frequently. To improve the detection of keypoints for this application, we developed a novel method that refines keypoint estimates by rotating the input videos. We select the best rotation for every frame with a graph-based global solver. Thereby, we improve keypoint detection of an arbitrary pose estimation algorithm, in particular for ‘hard’ keypoints. In the current proof-of-concept study, we show that our approach outperforms standard keypoint detection results in all categories and in all metrics, in injury-related out-of-balance and fall situations by a large margin as well as previous methods, in performance and robustness. The Injury Ski II dataset was made publicly available, aiming to facilitate the investigation of sports accidents based on computer vision in the future.
... The development of digital assistive systems for precise human motion evaluation has posed a formidable challenge in the field of artificial intelligence (AI) research [5]. This challenge is amplified by the diverse range of movements exhibited by individuals, variations in physical capabilities, and the inherent subjectivity involved in the algorithmic approaches used to assess human motion. ...
The demand for automated systems monitoring and supporting patients in their home-based recovery programs is substantial. While emerging technologies have been proposed as potential solutions to enhance at-home patient care, limited systems are in place due to their challenges in offering real-time monitoring and corrective feedback. Most proposed methodologies provide an overall measure or score for the executed movement. The proposed work involves the adaptation of an existing published dataset to cater to the needs of a system capable of remotely assisting patients, effectively acting as a virtual physical therapist able to provide corrective feedback. A dataset containing a set of three simple exercises for shoulder rehabilitation was processed. Each movement was meticulously annotated for temporal and cate- gorical motion domains to monitor exercise execution in terms of the range of motion completeness and to evaluate compensatory movement patterns. This work carries substantial significance by offering a standardized and easily accessible model for human movement data, thus fostering the advancement of digital assistive systems designed to support home-based rehabilitation programs.
... It is a method that drives autonomous learning on different devices by constantly processing data. Automating performance enables the ability to quickly respond to numerous adversities, such as improving characteristic combat positions and movements [3,4]. ...
This paper aims to demonstrate how design and digital media can have a relevant contribution to the improvement of Taekwondo athletes’ performance. This study focuses on answering the existing gap of a solution that allows quick and accurate access to data about the performance of martial arts athletes. This access to complex information, previously inaccessible or indecipherable to athletes and coaches, allowed, through digital design, the improvement of communication and a more personalized training feedback. The methodology developed was based on design thinking, in a work process that consisted of user identification, and the conception of a prototype in the user-centred design framework. The results obtained in the usability tests performed with Taekwondo athletes and coaches were demonstrative of the efficiency of the designed solution. These scores are also a stimulus for the potential replication and adaptation of the study in other martial arts.
... These findings regarding differences between winning and losing teams are based entirely on quantitative analysis of expert-based manual annotation of handball matches. However, obtaining such data is extremely costly (Carling et al., 2008;Cust et al., 2019). Objective and reliable labels require multiple domain experts to observe hours of video footage (O'Donoghue, 2007). ...
... The fusion of human acumen with algorithmic processing is shaping a more holistic and precise comprehension of athletic performance, breaking free from the constraints of conventional retrospective dissection. This new frontier empowers coaches and analysts to chart a course that transcends the limitations of the past, fostering innovation and yielding insights that might otherwise have eluded even the most seasoned experts (Cust et al. 2019). ...
Tennis has gained global popularity, prompting a surge in interest towards 3D video-based tennis motion recognition. Early action recognition, which predates activity completion, is a critical classification task to preempt adverse outcomes. Prior research emphasizes effective feature extraction and modeling for swift, accurate classification, despite limited data availability. To establish a robust foundation, this study introduces an anticipatory action prediction module preceding the recognition component. The module forecasts subsequent motions based on observed ones, using an LSTM-GAN structure to mitigate motion blurring and generate predictions. This paper presents an innovative framework that leverages deep learning, particularly dilated neural networks, for real-time spatio-temporal tennis analysis on standard hardware, aiming to enhance player performance insights and action prediction through TensorFlow. The dilated RNN and CNN are integrated into the recognition module for comprehensive spatiotemporal feature modeling. To foster synergy between the prediction and recognition modules, a hard class mining mechanism is devised to enhance the learning capabilities of challenging class samples. As a result, the LSTM architecture combined with GAN provides an excellent 92.1 Precision, 91.2 Recall, 94.5 F-1 score and 95.0 Accuracy in action recognition and prediction of tennis sport, which is significantly higher than classical models i.e. GAN, Conv3DJ, Co-occurrence LSTM, and GAN + L1 + Mining.
... There was a 2018 roundtable discussion [177] about the latest findings in RGB-depth-based HAR can be seen in FIGURE 10. Using features from both the data streams and the activity recognition, another study characterized vision-based HAR [178] AI-based models for HAR are illustrated in FIGURE 14, and several DL models are used for HAR using EMG signals. Here are some models that were commonly used for HAR. ...
Human activity recognition (HAR) has become increasingly popular in recent years due to its potential to meet the growing needs of various industries. Electromyography (EMG) is essential in various clinical and biological settings. It is a metric that helps doctors diagnose conditions that affect muscle activation patterns and monitor patients’ progress in rehabilitation. Despite its widespread Application, existing methods for recording and interpreting EMG data need more signal detection and robust categorization. Recent material science and Artificial Intelligence (AI) developments have significantly improved EMG detection. With an increasingly elderly patient population, HAR is increasingly used to monitor patients’ Activities of Daily Living (ADLs) in healthcare settings. It is also being used in security settings to identify suspect behavior, and Surface EMG (sEMG) is a potential non-invasive treatment for HAR since it monitors muscle contractions during exercise. sEMG and AI have revolutionized HAR systems in recent years. Sophisticated methods are required to recognize, break down, manufacture, and classify the EMG signals obtained by muscles. This review summarizes the various research papers based on HAR with EMG. AI has made tremendous contributions to biomedical signals classification. The different approaches of preprocessing, feature extraction, Reduction techniques, Deep Learning (DL) and Machine Learning (ML) based classification methods of EMG signals are then briefly explained. We focused on latest ML/DL methods used in HAR, Hardware involved in HAR with EMG and EMG based Application. We also discovered open issues and future research direction that may point to new lines of inquiry for ongoing research toward EMG-based detection.
... Similarly, Kautz et al. [17] evaluated the application of deep neural networks for the recognition of volleyball-specific movement data collected by a tri-axial wearable sensor. In addition, Cust et al. [18] provides a systematic review of 52 studies on the topic of sport-specific movement recognition using machine and deep learning. ...
Emerging smart devices have gathered increasing popularity within the sports community, presenting a promising avenue for enhancing athletic performance. Among these, the Rise Dynamics Alpha (RD 𝛼
) smart gloves exemplify a system designed to quantify boxing techniques. The objective of this study is to expand upon the existing RD 𝛼
system by integrating machine-learning models for striking technique and target object classification, subsequently validating the outcomes through empirical analysis. For the implementation, a data-acquisition experiment is conducted based on which the most common supervised ML models are trained: decision tree, random forest, support vector machine, k-nearest neighbor, naive Bayes, perceptron, multi-layer perceptron, and logistic regression. Using model optimization and significance testing, the best-performing classifier, i.e., support vector classifier (SVC), is selected. For an independent evaluation, a final experiment is conducted with participants unknown to the developed models. The accuracy results of the data-acquisition group are 93.03% (striking technique) and 98.26% (target object) and for the independent evaluation group 89.55% (striking technique) and 75.97% (target object). Therefore, it is concluded that the system based on SVC is suitable for target object and technique classification.
... Recent analytic advances can be found in the domain of machine learning, which can generally be described as computer systems that learn and adapt without following specific instructions. One example is computer vision, which contains models that can learn from visual data to automatically detect and classify sport-specific movements (Cust et al., 2019). In general, the field invented and re-discovered a plethora of statistical models, many of which are promising because the models are distribution-free and are able to find complex relationships in data. ...
Selecting the right individuals for a sports team, organization, or military unit has a large influence on the achievements of the organization. However, the approaches commonly used for selection are either not reporting predictive performance or not explainable (i.e., black box models). In the present study, we introduce a novel approach to selection research, using various machine learning models. We examined 296 recruits, of whom 214 dropped out, who performed a set of physical and psychological tests. On this data, we compared four machine learning models on their predictive performance, explainability, and stability. We found that a stable rule-based (SIRUS) model was most suitable for classifying dropouts from the special forces selection program. With an averaged area under the curve score of 0.75, this model had a high predictive performance, and was most explainable and stable compared to the alternative models. Furthermore, we found that both physical and psychological variables were related to dropout. More specifically, a lower score on the 2800 meters time, sprint time, connectedness, skin folds, and fear of failure were most strongly associated with graduation. We discuss how researchers and practitioners can benefit from these insights.
... The processing and classification of motion sensor signals follow similar stages as in BCI and EMG systems, including preprocessing, feature extraction, feature selection, and classification [111][112][113]. Here, we provide a concise summary while emphasizing the methods specific to motion sensors: ...
... Nonetheless, this had also been pointed as a limitation due to non-ecological environments by the simulation of circuits or matches. 24 Besides, such devices are expensive and not all teams have access to them. The lack of analysis of accelerometer-based variables is another limitation that may provide further state-of-the-art knowledge, considering that accelerating or decelerating with or without changing direction has been reported as very important in soccer. ...
This study aimed to compare the agreement of total distance (TD), high-speed running (HSR) distance, and sprint distance during 16 official soccer matches between a global navigation satellite system (GNSS) and an optical-tracking system. A total of 24 male soccer players, who are actively participating in the Polish Ekstraklasa professional league, were included in the analysis conducted during official competitions. The players were systematically monitored using Catapult GNSS (10-Hz, S7) and Tracab optical-tracking system (25-Hz, ChyronHego). TD, HSR distance, sprint distance, HSR count (HSRC), and sprint count (SC) were collected. The data were extracted in 5-min epochs. A statistical approach was employed to visually examine the relationship between the systems based on the same measure. Additionally, R2 was utilized as a metric to quantify the proportion of variance accounted for by a variable. To assess agreement, Bland-Altman plots were visually inspected. The data from both systems were compared using the estimates derived from the intraclass correlation (ICC) test and Pearson product-moment correlation. Finally, a paired t-test was employed to compare the measurements obtained from both systems. The interaction between Catapult and Tracab systems revealed an R2 of 0.717 for TD, 0.512 for HSR distance, 0.647 for sprint distance, 0.349 for HSRC, and 0.261 for SC. The ICC values for absolute agreement between the systems were excellent for TD (ICC = 0.974) and good for HSR distance (ICC = 0.766), sprint distance (ICC = 0.822). The ICC values were not good for HSRCs (ICC = 0.659) and SCs (ICC = 0.640). t-test revealed significant differences between Catapult and Tracab for TD (p < 0.001; d = -0.084), HSR distance (p < 0.001; d = -0.481), sprint distance (p < 0.001; d = -0.513), HSRC (p < 0.001; d = -0.558), and SC (p < 0.001; d = -0.334). Although both systems present acceptable agreement in TD, they may not be perfectly interchangeable, which sports scientists and coaches must consider when using them.
... In recent years, there has been a notable increase in comprehensive surveys exploring the applications of machine learning and deep learning in sports performance. These surveys cover a wide range of topics, including the recognition of sportsspecific movements [5], mining sports data [6], and employing AI techniques in team sports [7]. While some surveys focus on specific sports like soccer [7] and badminton [8], others concentrate on particular tasks within computer vision, such as video action recognition [9], video action quality assessment [10], and ball tracking [11]. ...
Deep learning has the potential to revolutionize sports performance, with applications ranging from perception and comprehension to decision. This paper presents a comprehensive survey of deep learning in sports performance, focusing on three main aspects: algorithms, datasets and virtual environments, and challenges. Firstly, we discuss the hierarchical structure of deep learning algorithms in sports performance which includes perception, comprehension and decision while comparing their strengths and weaknesses. Secondly, we list widely used existing datasets in sports and highlight their characteristics and limitations. Finally, we summarize current challenges and point out future trends of deep learning in sports. Our survey provides valuable reference material for researchers interested in deep learning in sports applications.
... Unfortunately, adopting conventional ML-based motion classifiers for on-sensor motion recognition is challenging due to four shortcomings. The Compute Constraints of Integrated Processors: A microcontroller used for tinyML applications typically has 1-2 MB of SRAM and flash [23], while an on-sensor processor typically has 8 kB of SRAM and 32 kB of volatile data memory (no flash) 5 . Furthermore, the architecture of onsensor processors (e.g., STRED) is different from generalpurpose microcontrollers (e.g., ARM Cortex M4), making tinyML compiler suites incompatible for use on-sensor. ...
Inertial sensors provide a low-power and high-fidelity pathway for state estimation and sensor fusion. Inertial measurement units now feature on-chip processors for ultra-low-power information fusion, signal processing, and neural network-based classification at the extreme edge. However, accounting for domain shifts, personalized inference requirements, and application diversity makes adopting existing learning-enabled on-device training, classification, and fusion frameworks for on-sensor processors difficult. This paper introduces a method for personalized and on-device learning for on-chip classification, inference, and information fusion applications. The proposed framework automatically segments and stores quantized gravity vector image templates and axes variance information of motion artifacts during training. During inference, templates created from the time-series windows are matched against uniform blurred templates using the universal image quality index. An adaptive rep counting module robust to varying motion primitives counts repetitions of matched motion primitives. The framework requires no human-engineered parameters and allows for the personalization and addition of new motion artifacts. Our framework recognizes human activities with 96.7% test accuracy and achieves an average rep count error of 0.44, while reducing the memory usage by 1000-2000× over existing tiny machine learning on-device learning techniques, allowing on-sensor learning and inference under 8 KB of memory.
... While there are several deep learning methods, previous findings that indicate a convolutional neural network (CNN) trained with ECG or accelerometry time series data can accurately classify patients with/without cardiac dysfunction [41] and different human activities [37,39,40], respectively. Notable advantages of CNNs include computational efficiency and little to no manual feature engineering [39,49]. For these reasons, a CNN appeared to be a good starting point to explore if deep learning could automatically identify features in ECG and/or accelerometry signals that differentiate between mTBI PSCs. Figure 1 shows the data pipeline from patient identification through the experiments executed in this study. ...
Although injury mechanisms of mild traumatic brain injury (mTBI) may be similar across patients, it is becoming increasingly clear that patients cannot be treated as one homogenous group. Several predominant symptom clusters (PSC) have been identified, each requiring specific and individualised treatment plans. However, objective methods to support these clinical decisions are lacking. This pilot study explored whether wearable sensor data collected during the Buffalo Concussion Treadmill Test (BCTT) combined with a deep learning approach could accurately classify mTBI patients with physiological PSC versus vestibulo-ocular PSC. A cross-sectional design evaluated a convolutional neural network model trained with electrocardiography (ECG) and accelerometry data. With a leave-one-out approach, this model classified 11 of 12 (92%) patients with physiological PSC and 3 of 5 (60%) patients with vestibulo-ocular PSC. The same classification accuracy was observed in a model only using accelerometry data. Our pilot results suggest that adding wearable sensors during clinical tests like the BCTT, combined with deep learning models, may have the utility to assist management decisions for mTBI patients in the future. We reiterate that more validation is needed to replicate the current results.
... Therefore, inertial measurement units (IMU) have been utilized. These sensors are considered non-obstructive [10] (e.g., Catapult Sports Vector S7 weights 53 g with dimensions of 81 × 43 × 16 mm) and have been used by both outdoor and indoor team sports to measure physical demands [1,4,[11][12][13]. ...
In sports like football, knowledge concerning physical demands is used for training sessions and matches. Inertial measurement units are often used to estimate physical demands. Catapult Sports has recently developed, based on acceleration and gyroscope data, the Football Movement Profile algorithm that enables to categorize physical demands into locomotion categories facilitating the interpretation of practitioners. The aim of this study was to assess the validity of the locomotion predictions derived from the Football Movement Profile during controlled drills. Data were collected from 41 elite youth football players executing controlled drills replicating linear locomotion at five velocities, i.e., 5.4, 12, 18, 19.8, and 21.6 km/h and non-linear locomotion. An overall good agreement (> 95%) was found for lower velocity linear movements (≤ 12 km/h), but the agreement decreased with increasing velocities (67% at 21.6 km/h). A poor agreement was found for the lower velocity non-linear movements (78%) but increased with higher velocities (> 98%). The relationship between the magnitude and type of classification errors among the players suggests that it would be beneficial to establish individual calibration thresholds for the algorithm.
... Distance estimation based on magneto-inertial measurement units (MIMUs) can certainly rely on laws of motion and biomechanical models, as completed for height of countermovement jumps [24,25], but the poor quality of the available signals [23] has led to exploiting biomechanical features obtained from MIMUs as input to ad-hoc machine learning (ML) models in several sport applications [26][27][28]. Few ML approaches used SP data to assess jump-related variables only for the countermovement jump (CMJ): jump height [29], jump power [30,31], and fatigue [32]. ...
The length of the standing long jump (SLJ) is widely recognized as an indicator of developmental motor competence or sports conditional performance. This work aims at defining a methodology to allow athletes/coaches to easily measure it using the inertial measurement units embedded on a smartphone. A sample group of 114 trained young participants was recruited and asked to perform the instrumented SLJ task. A set of features was identified based on biomechanical knowledge, then Lasso regression allowed the identification of a subset of predictors of the SLJ length that was used as input of different optimized machine learning architectures. Results obtained from the use of the proposed configuration allow an estimate of the SLJ length with a Gaussian Process Regression model with a RMSE of 0.122 m in the test phase, Kendall’s τ < 0.1. The proposed models give homoscedastic results, meaning that the error of the models does not depend on the estimated quantity. This study proved the feasibility of using low-cost smartphone sensors to provide an automatic and objective estimate of SLJ performance in ecological settings.
... Various activities have been studied using accelerometers including: activity of daily living [10,11] with the UCI50 and Wireless Sensor Data Mining (WISDM) datasets [12]; factory workers [13]; food preparation [14]; tennis; snowboarding; weight lifting; rugby and running [15,16]. Johnson et al. [17] used cameras and load cells with deep learning to predict ground reaction forces. ...
... Therefore, the accurate identification of cross-country skiing technical movements can effectively help athletes to optimize their gliding strategies and improve their technical movements, thus improving their performance. There are three main methods for cross-country skiing technical movement recognition [2]: one is based on improving athletes' equipment [3,4], but cross-country skiing has high demands on equipment and professional athletes need a long time to adapt to the modified equipment, which limits the promotion and application of this method. Another is based on wearable micro-sensors [3,5,6], such as Global Navigation Satellite System (GNSS) data and seven Inertial Measurement Units (IMUs), which are used to analyze different cycle characteristics and sub-technology types during skiing. ...
Technical motion recognition in cross-country skiing can effectively help athletes to improve their skiing movements and optimize their skiing strategies. The non-contact acquisition method of the visual sensor has a bright future in ski training. The changing posture of the athletes, the environment of the ski resort, and the limited field of view have posed great challenges for motion recognition. To improve the applicability of monocular optical sensor-based motion recognition in skiing, we propose a monocular posture detection method based on cooperative detection and feature extraction. Our method uses four feature layers of different sizes to simultaneously detect human posture and key points and takes the position deviation loss and rotation compensation loss of key points as the loss function to implement the three-dimensional estimation of key points. Then, according to the typical characteristics of cross-country skiing movement stages and major sub-movements, the key points are divided and the features are extracted to implement the ski movement recognition. The experimental results show that our method is 90% accurate for cross-country skiing movements, which is equivalent to the recognition method based on wearable sensors. Therefore, our algorithm has application value in the scientific training of cross-country skiing.
... It may also provide a way to distinguish between the eccentric and concentric COD phases. Additionally, automated pattern recognition algorithms could be implemented to recognize the signal shape for each COD angle (90°, 180°, etc.) to enable COD analysis during match play 41 . For detecting the end of the T-test, the validity of the GNSS ground speed should be inspected due to its low sampling frequency (10 Hz). ...
The Agility T-test is a standardized method to measure the change-of-direction (COD) ability of athletes in the field. It is traditionally scored based on the total completion time, which does not provide information on the different CODs. Augmenting the T-test with wearable sensors provides the opportunity to explore new metrics. Towards this, data of 23 professional soccer players were recorded with a trunk-worn GNSS-IMU (Global Navigation Satellite System-Inertial Measurement Unit) device. A method for detecting the four CODs based on the wavelet-denoised antero-posterior acceleration signal was developed and validated using video data (60 Hz). Following this, completion time was estimated using GNSS ground speed and validated with the photocell data. The proposed method yields an error (mean ± standard deviation) of 0 ± 66 ms for the COD detection, − 0.16 ± 0.22 s for completion time, and a relative error for each COD duration and each sequential movement durations of less than 3.5 ± 16% and 7 ± 7%, respectively. The presented algorithm can highlight the asymmetric performance between the phases and CODs in the right and left direction. By providing a more comprehensive analysis in the field, this work can enable coaches to develop more personalized training and rehabilitation programs.
... Auch wenn die Anwendung von KI und ML in der Trainingswissenschaft noch in den Kinderschuhen steckt, zeigen sich erste Anwendungsfelder. So können mit Sensordaten gespeiste ML Algorithmen in der Bewegungsanalyse verschiedenster Sportarten spezifische Bewegungen automatisiert erkennen(Cust et al. 2019;Jang et al. 2018). Die Rückmeldung über qualitative oder quantitative Aspekte der Bewegung ermöglicht dann ein spezifisches Training an Defiziten in der Bewegungsausführung oder kann zur Belastungssteuerung genutzt werden. ...
Neueste miniaturisierte Sensortechnik und deren Applikationen beeinflussen zunehmend gesellschaftliche und soziale Lebenswelten und kommen in Sport und Gesundheitswesen verstärkt zur Anwendung. Die Aufzeichnung, Verarbeitung und Darstellung verschiedenster biophysikalischer Signale und Umweltfaktoren im Alltag, im Training, bei Wettkämpfen sowie zur Regeneration stellen eine große Herausforderung im Hinblick auf die Informationsnutzung für Training, Gesundheit und aktives Verhalten dar. Da kommerziell vermarktete Wearables in Deutschland derzeit nicht reguliert und keiner unabhängigen Überprüfung der Reliabilität und Validität unterliegen, werden zukünftig Experten gebraucht, welche die Qualität der zur Verfügung gestellten Daten bewerten können. „Zukünftige“ Sportwissenschaftler müssen ein Verständnis für die jeweilige Sportart, aber auch über die zur Verfügung stehenden Daten sowie deren Analysen, Interpretation und Darstellung entwickeln.
... This knowledge quickly reached the field of sports medicine, with contributions in the field of genomics, sportomics, biomechanics, kinesiology, sports training and performance, among others (Cust et al., 2019;Tanisawa et al., 2020;Antink et al., 2021;Gonçalves et al., 2022). ...
The present study aimed to investigate the possible correlations between the cytokine and adipokine Tumour Necrosis Factor Alpha with parameters of body composition and lipid metabolism in young, high-level athletes after an incremental treadmill test observed in a sample of five individuals, male, high-level running athletes who the difficulty of treating large databases with different individuals, multiple biomarkers, and collection times, in addition to physical parameters and sample characteristics, added to the decrease in new findings induced by the application of statistical tools of univariate analysis, indicate the need to apply exploratory machine learning strategies, generating holistic and integrated analysis of the results. The present study showed a negative correlation between TNF and HDL and a similarity between the same TNF and LDL. These findings do not indicate a cause-and-effect relationship but suggest a possible modulation of the immune system, lipid metabolism, and exercise that requires further investigation.
... Features must be chosen as appropriate for the context they are describing: those biomechanically linked with the jump (5) would be reasonable candidates for obtaining a valid estimate; but it is also essential to embed features that can be predictive, and thus possibly compensate for, the wobbling oscillations mentioned above. This is in line with recent trends having data science emerging as a discipline capable of supporting findings related to sports-related issues through the use of automated methods (46)(47)(48). ...
Introduction: The peak height reached in a countermovement jump is a well established performance parameter. Its estimate is often entrusted to force platforms or body-worn inertial sensors. To date, smartphones may possibly be used as an alternative for estimating jump height, since they natively embed inertial sensors.
Methods: For this purpose, 43 participants performed 4 countermovement jumps (172 in total) on two force platforms (gold standard). While jumping, participants held a smartphone in their hands, whose inertial sensor measures were recorded. After peak height was computed for both instrumentations, twenty-nine features were
extracted, related to jump biomechanics and to signal time-frequency characteristics, as potential descriptors of soft tissues or involuntary arm swing artifacts. A training set (129 jumps – 75%) was created by randomly selecting elements from the initial dataset, the remaining ones being assigned to the test set (43 jumps – 25%). On the training set only, a Lasso regularization was applied to reduce the number of features, avoiding possible multicollinearity. A multi-layer perceptron with one hidden layer was trained for estimating the jump height from the reduced feature set. Hyperparameters optimization was performed on the multi-layer perceptron using a grid search approach with 5-fold cross validation. The best model was chosen according to the minimum negative mean absolute error.
Results: The multi-layer perceptron greatly improved the accuracy (4 cm) and precision (4 cm) of the estimates on the test set with respect to the raw smartphone measures estimates (18 and 16 cm, respectively). Permutation feature importance was performed on the trained model in order to establish the influence that each feature had on the outcome. The peak acceleration and the braking phase duration resulted the most influential features in the final model. Despite not being accurate enough, the height computed through raw smartphone measures was still among the most influential features.
Discussion: The study, implementing a smartphone-based method for jump height
estimates, paves the way to method release to a broader audience, pursuing a
democratization attempt.
... The model can give coaching and medical staff tackle-specific measurements, in real-time, which can be used in injury prevention and rehabilitation strategies. Following on from this, (Cust et al. 2018) reviewed the ways that machine learning and AI can be used to classify certain movements in sport. ...
The Sports Analytics Market is growing rapidly, in 2020 it was valued at over $1 billion and is expected to reach over $5 billion by 2026. However, even with this level of growth the use of Artificial Intelligence (AI) techniques have yet to fully be explored. The sports analytics domain presents a number of significant computational challenges for AI and Machine Learning. In this thesis, we propose a number novel methods for analysing team sports data to help sports teams utilise AI to improve their strategic and tactical decision making. By doing so, we present a number of contributions to the AI and sports analytics communities. In particular, we present a model for the tactical decisions that are made in football and show how game theoretic techniques can be used to optimise these. We focus on both the short-term decisions made for individual games, as well as longer term decisions to maximise performance over a season. We show that we can increase a teams chances of winning individual games by 16.1% and can increase a teams mean expected finishing position by up to 35.6%. We also, introduce a new model for valuing the teamwork between players in sports teams by assessing the outcomes of chains of interactions between the players in a team. We then present a novel model for forming teams based on this value and maximise teamwork by assessing the overlapping pairs in a team. Our model is shown to better predict the real-world performance of teams by up to 46% compared to models that ignore inter-agent interactions. Finally, we show how we can use natural language processing techniques to improve the traditional statistical methods for prediction sports match outcomes. We use domain expert written articles from the media to train our models and we show that by incorporating the features learned from the text, we can boost the accuracy of the traditional statistical methods by 6.9%.
... One option for developing the evidence-based training is to apply innovative technologies, such as wearable systems and machine learning in coaching practice. Such an approach would help the elite athletes to activate their potential capabilities and stretch human biological boundaries [11][12][13]. ...
Developing real-time biomechanical feedback systems for in-field applications will transfer human motor skills’ learning/training from subjective (experience-based) to objective (science-based). The translation will greatly improve the efficiency of human motor skills’ learning and training. Such a translation is especially indispensable for the hammer-throw training which still relies on coaches’ experience/observation and has not seen a new world record since 1986. Therefore, we developed a wearable wireless sensor system combining with artificial intelligence for real-time biomechanical feedback training in hammer throw. A framework was devised for developing such practical wearable systems. A printed circuit board was designed to miniaturize the size of the wearable device, where an Arduino microcontroller, an XBee wireless communication module, an embedded load cell and two micro inertial measurement units (IMUs) could be inserted/connected onto the board. The load cell was for measuring the wire tension, while the two IMUs were for determining the vertical displacements of the wrists and the hip. After calibration, the device returned a mean relative error of 0.87% for the load cell and the accuracy of 6% for the IMUs. Further, two deep neural network models were built to estimate selected joint angles of upper and lower limbs related to limb coordination based on the IMUs’ measurements. The estimation errors for both models were within an acceptable range, i.e., approximately ±12° and ±4°, respectively, demonstrating strong correlation existed between the limb coordination and the IMUs’ measurements. The results of the current study suggest a remarkable novelty: the difficulty-to-measure human motor skills, especially in those sports involving high speed and complex motor skills, can be tracked by wearable sensors with neglect movement constraints to the athletes. Therefore, the application of artificial intelligence in a wearable system has shown great potential of establishing real-time biomechanical feedback training in various sports. To our best knowledge, this is the first practical research of combing wearables and machine learning to provide biomechanical feedback in hammer throw. Hopefully, more wearable biomechanical feedback systems integrating artificial intelligence would be developed in the future.
... Workers in the construction industry are often exposed to physically demanding manual tasks with a high degree of ergonomic risk [71,72]. The rapid development of motion sensors in the construction industry enables proactive accident prevention by reducing the number of dangerous actions that commonly occur [73]. The authors of the articles Experience, Productivity, and Musculoskeletal Injury among Masonry Workers [74], Data Fusion of [71] want to achieve consistent results. ...
Due to the increasingly high proportion of manual activities in production processes, there is a constant risk of musculoskeletal disorders or work-related injuries. The risk of these problems is exacerbated by the trend towards an ageing working population. European legislation is pressing for improved working conditions to eliminate the risks associated with health problems for workers. For this reason, the application of ergonomics in this field is growing. Musculoskeletal disorders, which are most often caused by inappropriate working postures, are a major problem. There are many methods for evaluating working postures. However, there is a high degree of subjectivity in the risk assessment. Motion capture kinematic suits can ensure the objectivity of the assessment. This article discusses research on ergonomics assessment using motion capture technology. A systematic literature search method was used for the research, beginning with the determination of the research procedure, through the definition of the research queries, to the formulation of the research itself to identify relevant sources. The study presents the most widely used methods for assessing the ergonomics of work positions using motion capture technology, their advantages, and disadvantages. It also follows the trend in the number of publications between 2010 and 2022 in countries where the topic is most frequently addressed and in the industries where motion capture technology is used for ergonomics assessment in general. The research showed that this approach is most often used in industry and logistics, and less frequently in healthcare and sport. The authors agree that the most frequently used ergonomics assessment methods are not complex enough to be used in combination with motion capture and that a combination of the two is needed. At the same time, this technology has become very important in the field of ergonomic evaluation of work positions, offering a higher degree of objectivity, or can be combined with the use of virtual reality, but the evaluation systems are still not error-free and there is a need for continuous improvement.
... Machine learning and deep learning are the new technology in the computer world. According to (Cust et al., 2019) indicate the positive affect of using machine learning and deep learning to enhance the sport activity. It is mean sport activity can enhance and practices with using the technology. ...
Objectives:Sports participation has an important role to play. Sports activities are change environment to another. Also, using information and communication technology such as distance learning has become a rapidly growing trend. Online delivery platform becomes very important methods for the graduate and undergraduate students to achieves the advanced degrees and knowledge.The aim of this study to investigate the impact of online learning on Sports participation during pandemic (COVID-19). Methods: This study usedthe previous studiesto investigate theuse of online learning on Sports participation during pandemic (COVID-19). This research uses different databased such as Science direct, Google Scholar, Scopus and other databases. Results: This study comeswith some obstacles that face student in Saudi Arabia to practise the sport activities during the pandemic (COVID-19). Those obstacles should consider by Saudi ministry of education.
We are currently witnessing an unprecedented era of digital transformation in sports, driven by the revolutions in Artificial Intelligence (AI), Virtual Reality (VR), Augmented Reality (AR), and Data Visualization (DV). These technologies hold the promise of redefining sports performance analysis, automating data collection, creating immersive training environments, and enhancing decision-making processes. Traditionally, performance analysis in sports relied on manual data collection, subjective observations, and standard statistical models. These methods, while effective, had limitations in terms of time and subjectivity. However, recent advances in technology have ush-ered in a new era of objective and real-time performance analysis. AI has revolutionized sports analysis by streamlining data collection, processing vast datasets, and automating information synthesis. VR introduces highly realistic training environments, allowing athletes to train and refine their skills in controlled settings. AR overlays digital information onto the real sports environment, providing real-time feedback and facilitating tactical planning. DV techniques convert complex data into visual representations, improving the understanding of performance metrics. In this paper, we explore the potential of these emerging technologies to transform sports performance analysis, offering valuable resources to coaches and athletes. We aim to enhance athletes' performance, optimize training strategies, and inform decision-making processes. Additionally, we identify challenges and propose solutions for integrating these technologies into current sports analysis practices. This narrative review provides a comprehensive analysis of the historical context and evolution of performance analysis in sports science, highlighting current methods' merits and limitations. It delves into the transformative potential of AI, VR, AR, and DV, offering insights into how these tools can be integrated into a theoretical model.
This article focuses on evaluating the efficacy of intelligent image processing techniques using deep learning algorithms in the context of football, to present pragmatic solutions for enhancing the functional strength training of football players. The article commences by delving into the prevailing research landscape concerning image recognition in football. It then embarks on a comprehensive examination of the prevailing landscape in soccer image recognition research. Subsequently, a novel soccer image classification model is meticulously crafted through the fusion of Space-Time Graph Neural Network (STGNN) and Bi-directional Long Short-Term Memory (BiLSTM). The devised model introduces the potency of STGNN to extract spatial features from sequences of images, adeptly harnessing spatial information through judiciously integrated graph convolutional layers. These layers are further bolstered by the infusion of graph attention modules and channel attention modules, working in tandem to amplify salient information within distinct channels. Concurrently, the temporal dimension is adroitly addressed by the incorporation of BiLSTM, effectively capturing the temporal dynamics inherent in image sequences. Rigorous simulation analyses are conducted to gauge the prowess of this model. The empirical outcomes resoundingly affirm the potency of the proposed deep hybrid attention network model in the realm of soccer image processing tasks. In the arena of action recognition and classification, this model emerges as a paragon of performance enhancement. Impressively, the model notched an accuracy of 94.34 %, precision of 92.35 %, recall of 90.44 %, and F1-score of 89.22 %. Further scrutiny of the model's image recognition capabilities unveils its proficiency in extracting comprehensive features and maintaining stable recognition performance when applied to football images. Consequently, the football intelligent image processing model based on deep hybrid attention networks, as formulated within this article, attains high recognition accuracy and demonstrates consistent recognition performance. These findings offer invaluable insights for injury prevention and personalized skill enhancement in the training of football players.
Background
Movement quality is typically assessed by drawing comparisons against predetermined movement standards. Movements are often discretely scored or labelled against pre-set criteria, though movement quality can also be evaluated using motion-related measurements (e.g., spatio-temporal parameters and kinematic variables). Wearable technology has the potential to measure and assess movement quality and offer valuable, practical feedback.
Objectives
A systematic approach was taken to examine the benefits associated with multi-sensor and multiple wearable-device usage, compared with unimodal applications, when assessing movement quality. Consequently, this review considers the additional variables and features that could be obtained through multi-sensor devices for use in movement analyses. Processing methods and applications of the various configurations were also explored.
Methods
Articles were included within this review if they were written in English, specifically studied the use of wearable sensors to assess movement quality, and were published between January 2010 and December 2022. Of the 62,635 articles initially identified, 27 papers were included in this review. The quality of included studies was determined using a modified Downs and Black checklist, with 24/27 high quality.
Results
Fifteen of the 27 included studies used a classification approach, 11 used a measurement approach, and one used both methods. Accelerometers featured in all 27 studies, in isolation (n = 5), with a gyroscope (n = 9), or with both a gyroscope and a magnetometer (n = 13). Sampling frequencies across all studies ranged from 50 to 200 Hz. The most common classification methods were traditional feature-based classifiers (n = 5) and support vector machines (SVM; n = 5). Sensor fusion featured in six of the 16 classification studies and nine of the 12 measurement studies, with the Madgwick algorithm most prevalent (n = 7).
Conclusions
This systematic review highlights the differences between the applications and processing methods associated with the use of unimodal and multi-sensor wearable devices when assessing movement quality. Further, the use of multiple devices appears to increase the feasibility of effectively assessing holistic movements, while multi-sensor devices offer the ability to obtain more output metrics
Training load (TL) is frequently documented among team sports and the development of emerging technology (ET) is displaying promising results towards player performance and injury risk identification. The aim of this systematic review was to identify ETs used in field-based sport to monitor TL for injury/performance prediction and provide sport specific recommendations by identifying new data generation in which coaches may consider when tracking players for an increased accuracy in training prescription and evaluation among field-based sports. Data was extracted from 60 articles following a systematic search of CINAHL, SPORTDiscus, Web of Science and IEEE XPLORE databases. Global positioning system (GPS) and accelerometers were common external TL tools and Rated Perceived Exertion (RPE) for internal TL. A collection of analytics tools were identified when investigating injury/performance prediction. Machine Learning showed promising results in many studies, identifying the strongest predictive variables and injury risk identification. Overall, a variety of TL monitoring tools and predictive analytics were utilized by researchers and were successful in predicting injury/performance, but no common method taken by researchers could be identified. This review highlights the positive effect of ETs, but further investigation is desired towards a ‘gold standard” predictive analytics tool for injury/performance prediction in field-based team sports.
In Cricket the environment is bound to undergo quick and precise motion that is time consuming and complex for analyzers and coaches to draw conclusions from and for audiences to follow. Though there are already existing methods utilizing bounding boxes for detection, their scope is limited considering the various biomechanical parameters and angles that are generally a critical source of information in this sport. This paper discusses the solution by using key point detection for angle estimation and orientation of the objects. Keywordsmovementsbiomechanical parametersangleskeypoints
Implementations of artificial intelligence and machine learning are becoming commonplace in multiple application domains. This is in part due to advancements in computing hardware that have helped outsource the computation of resource-intensive mathematics related to artificial intelligence and machine learning to the chips of multi-core and parallel computing architectures. Partly it is due to the widespread appeal of machine learning as a suite of handy tools to fix practical issues. Many fields have become beneficiaries of artificial intelligence and machine learning and cardiorespiratory rehabilitation is no exception.The aim of this paper is to review the current state of the art of the applications of artificial intelligence and machine learning in cardiorespiratory rehabilitation. We have taken a multidimensional view to addressing the needs and utility of artificial intelligence and machine learning in cardiorespiratory rehabilitation. We start with the most primitive applications of machine learning reported in existing literature in making medical devices for analyzing heartbeats and respiratory functions. We then discuss more recent approaches including deep learning to analyze performance or suggest alternative choices for food or exercise. Applications and utility of most recent feats such as explainable artificial intelligence are also discussed and conclusions around the current state of the art and possible future directions are proposed.
Human activity recognition (HAR) for boxing training typically requires specialized hardware or large labeled datasets to identify different types of punches and analyze their basic properties. However, measuring the unpredictability of boxers, a key characteristic to deceive opponents during combats, has remained an unexplored challenge. We tackled this challenge by combining a series of novel techniques and ideas. Unpredictability was computed as the entropy of the trajectories of the Markov chain characterizing the training session, estimated from the combinations (sequences of punches) thrown by the boxer to the training bag. Punch combinations were obtained by detecting punches as outliers in the raw acceleration sensor datastream coming from the bag, and separating them by type with an ensemble of PCA
$+$
GMM clusters. Punches were then divided into combinations using statistical analysis of separation intervals. The detection procedure works even with noisy data, in contrast to force-threshold methods commonly reported in the literature. As distinctive features of our system, it uses only unlabeled data from the current training session, and sensor positions and orientations inside the bag are unconstrained. However, punch clustering achieves up to 94% accuracy, comparable to state-of-the-art supervised approaches. These features are currently unseen in the literature and demonstrate the capacity for unsupervised learning techniques to address challenging problems in sports. Multiple validation experiments were consistent with boxers’ level and session performance. Overall, our approach provides significant contributions to the field of HAR and has the potential to improve the way boxing and other sports are trained and evaluated.
Inertial Measurement Units (IMU) and machine learning are strong tools in quantifying physical demands in sports, such as handball. However, the detection of both locomotion and throw events simultaneously has not been a topic for much investigation. Wherefore, the aim of this study was to publicise a method for training an extreme gradient boosting model capable of identifying low intensity, dynamic, running and throw events. Twelve adults with varying experience in handball wore an IMU on the back while being video recorded during a handball match. The video recordings were used for annotating the four events. Due to the small sample size, a leave-one-subject-out (LOSO) approach was conducted for the modelling and feature selection. The model had issues identifying dynamic movements (F1-score = 0.66 ± 0.07), whereas throw (F1-score = 0.95 ± 0.05), low intensity (F1-score = 0.93 ± 0.02) and running (F1-score = 0.86 ± 0.05) were easier to identify. Features such as IQR and first zero crossing for most of the kinematic characteristics were among the most important features for the model. Therefore, it is recommended for future research to look into these two features, while also using a LOSO approach to decrease likelihood of artificially high model performance.
Upper limb tennis injuries are primarily chronic, resulting from repetitive overuse. We developed a wearable device which simultaneously measures risk factors (grip strength, forearm muscle activity, and vibrational data) associated with elbow tendinopathy development resulting from tennis players’ technique. We tested the device on experienced (n = 18) and recreational (n = 22) tennis players hitting forehand cross-court at both flat and topspin spin levels under realistic playing conditions. Using statistical parametric mapping analysis, our results showed that all players showed a similar level of grip strength at impact, regardless of spin level, and the grip strength at impact did not influence the percentage of impact shock transfer to the wrist and elbow. Experienced players hitting with topspin exhibited the highest ball spin rotation, low-to-high swing path brushing action, and shock transfer to the wrist and elbow compared to the results obtained while hitting the ball flat, or when compared to the results obtained from recreational players. Recreational players exhibited significantly higher extensor activity during most of the follow through phase compared to the experienced players for both spin levels, potentially putting them at greater risk for developing lateral elbow tendinopathy. We successfully demonstrated that wearable technologies can be used to measure risk factors associated with elbow injury development in tennis players under realistic playing conditions.
Balance is a common performance but nevertheless an essential part of performance analysis investigations in ski. Many skier pay attention to the training of balance ability in training. Inertial Measurement Unit, as a kind of Multiplex-type human motion capture system, is widely used because of its humanized human-computer interaction design, low energy consumption and more freedom provided by the environment. The purpose of this research is to use sensor to establish a kinematics dataset of balance test tasks extracted from skis to help quantify skier’ balance ability. Perception Neuron Studio motion capture device is used in present. The dataset contains a total of 20 participants’ data (half male) of the motion and sensor data, which is collected at a 100 Hz sampling frequency. To our knowledge, this dataset is the only one that uses a BOSU ball in the balance test. We hope that this dataset will contribute to multiple fields of cross-technology integration in physical training and functional testing, including big-data analysis, sports equipment design and sports biomechanical analysis.
To understand human behaviors, action recognition based on videos is a common approach. Compared with image-based action recognition, videos provide much more information, reducing the ambiguity of actions. In the last decade, many works focus on datasets, novel models and learning approaches have improved video action recognition to a higher level. However, there are challenges and unsolved problems, in particular in sports analytics where data collection and labeling are more sophisticated, requiring people with domain knowledge and even sport professionals to annotate data. In addition, the actions could be extremely fast and it becomes difficult to recognize them. Moreover, in team sports like football and basketball, one action could involve multiple players, and to correctly recognize them, we need to analyze all players, which is relatively complicated. In this paper, we present a survey on video action recognition for sports analytics. We introduce more than ten types of sports, including team sports, such as football, basketball, volleyball, hockey and individual sports, such as figure skating, gymnastics, table tennis, tennis, diving and badminton. Then we compare numerous existing frameworks for sports analysis to present status quo of video action recognition in both team sports and individual sports. Finally, we discuss the challenges and unsolved problems in this area and to facilitate sports analytics, we develop a toolbox using PaddlePaddle, which supports football, basketball, table tennis and figure skating action recognition.
Artificial intelligence and its subcategories of machine learning and deep learning are gaining increasing importance and attention in the context of sports research. This has also meant that the number of corresponding publications has become complex and unmanageably large in human terms. In the current state of the research field, there is a lack of bibliometric analysis, which would prove useful for obtaining insights into the large amounts of available literature. Therefore, the present work aims to identify important research issues, elucidate the conceptual structure of the research field, and unpack the evolutionary trends and the direction of hot topics regarding key themes in the research field of artificial intelligence in sports. Using the Scopus database, 1215 documents (reviews and articles) were selected. Bibliometric analysis was performed using VOSviewer and bibliometrix R package. The main findings are as follows: (a) the literature and research interest concerning AI and its subcategories is growing exponentially; (b) the top 20 most cited works comprise 32.52% of the total citations; (c) the top 10 journals are responsible for 28.64% of all published documents; (d) strong collaborative relationships are present, along with small, isolated collaboration networks of individual institutions; (e) the three most productive countries are China, the USA, and Germany; (f) different research themes can be characterized using author keywords with current trend topics, e.g., in the fields of biomechanics, injury prevention or prediction, new algorithms, and learning approaches. AI research activities in the fields of sports pedagogy, sports sociology, and sports economics seem to have played a subordinate role thus far. Overall, the findings of this study expand knowledge on the research situation as well as the development of research topics regarding the use of artificial intelligence in sports, and may guide researchers to identify currently relevant topics and gaps in the research.
The use of wearable devices for player motion analysis is becoming a popular method to measure kinematic parameters associated with player techniques in racket sports. This systematic review focused on identifying the current applications of wearable technology for player motion analysis in racket sports (tennis, table tennis, badminton and squash) through two research questions: (1) What are the existing applications of wearable technology for player motion analysis in racket sports? (2) What data analysis methods are used to quantify and evaluate player motion? A comprehensive search of MEDLINE, EMBASE, SPORTDiscus, Scopus, Web of Science, and IEEE databases was undertaken following PRISMA reporting guidelines. Included studies must have only used external wearable technology mounted to either the player or the racket for the potential application of player motion analysis in a racket sport, tested the wearable technology under normal playing conditions and not only focused on detecting/classifying the player's stroke or activity movement. Of the 6616 articles found, a total of 15 studies met the inclusion and exclusion criteria. Tennis and table tennis were the most popular sports researched and inertial measurement units and electromyography sensors were the most common types of sensors used. This review found that wearable devices were mainly used to analyze (1) movement similarities and differences of players at different playing levels, (2) variability in racket, upper limb and joint movement patterns and (3) movement differences associated with different ball spin levels. These technologies can create a new level of accessible personalized training in these sports.
In recent years smart sport equipments have achieved great success in professional and amateur sports, as well as body sensory systems; now discovering interesting knowledge in the surge of data from those embedded sensors used in sports is necessary and the focus of our research. In this paper, we investigate golf swing data classification method based on deep convolutional neural network (deep CNN) fed with multi-sensor golf swing signals. Our smart golf club integrates two orthogonally affixed strain gage sensors, 3-axis accelerometer and 3-axis gyroscope, and collects real-world golf swing data from professional and amateur golf players. Furthermore we explore the performance of our well-trained CNN-based classifier and evaluate it on the real-world test set in terms of common indicators including accuracy, precision-recall, and F1-score. Experiments and corresponding results show that our CNN-based model can satisfy the requirement of accuracy of golf swing classification, and outperforms support vector machine (SVM) method.
Recent technological developments have led to the production of inexpensive, non-invasive, miniature magneto-inertial sensors, ideal for obtaining sport performance measures during training or competition. This systematic review evaluates current evidence and the future potential of their use in sport performance evaluation. Articles published in English (April 2017) were searched in Web-of-Science, Scopus, Pubmed, and Sport-Discus databases. A keyword search of titles, abstracts and keywords which included studies using accelerometers, gyroscopes and/or magnetometers to analyse sport motor-tasks performed by athletes (excluding risk of injury, physical activity, and energy expenditure) resulted in 2040 papers. Papers and reference list screening led to the selection of 286 studies and 23 reviews. Information on sport, motor-tasks, participants, device characteristics, sensor position and fixing, experimental setting and performance indicators was extracted. The selected papers dealt with motor capacity assessment (51 papers), technique analysis (163), activity classification (19), and physical demands assessment (61). Focus was placed mainly on elite and sub-elite athletes (59%) performing their sport in-field during training (62%) and competition (7%). Measuring movement outdoors created opportunities in winter sports (8%), water sports (16%), team sports (25%), and other outdoor activities (27%). Indications on the reliability of sensor-based performance indicators are provided, together with critical considerations and future trends.
Background:
Analysis of lower limb exercises is traditionally completed with four distinct methods (i) 3D motion capture; (ii) depth-camera based systems (iii) visual analysis from a qualified exercise professional; (iv) self-assessment. Each method is associated with a number of limitations.
Objective:
The aim of this systematic review is to synthesize and evaluate studies which have investigated the capacity for inertial measurement unit (IMU) technologies to assess movement quality in lower limb exercises.
Data Sources:
A systematic review of PubMed, ScienceDirect and Scopus was conducted.
Study Eligibility Criteria:
Articles written in English and published in the last 10 years which contained an IMU system for the analysis of repetition-based targeted lower limb exercises were included.
Study Appraisal and Synthesis Methods:
The quality of included studies was measured using an adapted version of the STROBE assessment criteria for cross-sectional studies. The studies were categorised in to three groupings: exercise detection, movement classification or measurement validation. Each study was then qualitatively summarised.
Results:
From the 2452 articles that were identified with the search strategies, 47 papers are included in this review.
Conclusions:
Wearable inertial sensor systems for analysing lower limb exercises are a rapidly growing technology. Research over the past ten years has predominantly focused on validating measurements that the systems produce and classifying users’ exercise quality. There have been very few user evaluation studies and no clinical trials in this field to date.
The automatic classification of sub-techniques in classical cross-country skiing provides unique possibilities for analyzing the biomechanical aspects of outdoor skiing. This is currently possible due to the miniaturization and flexibility of wearable inertial measurement units (IMUs) that allow researchers to bring the laboratory to the field. In this study, we aimed to optimize the accuracy of the automatic classification of classical cross-country skiing sub-techniques by using two IMUs attached to the skier’s arm and chest together with a machine learning algorithm. The novelty of our approach is the reliable detection of individual cycles using a gyroscope on the skier’s arm, while a neural network machine learning algorithm robustly classifies each cycle to a sub-technique using sensor data from an accelerometer on the chest. In this study, 24 datasets from 10 different participants were separated into the categories training-, validation- and test-data. Overall, we achieved a classification accuracy of 93.9% on the test-data. Furthermore, we illustrate how an accurate classification of sub-techniques can be combined with data from standard sports equipment including position, altitude, speed and heart rate measuring systems. Combining this information has the potential to provide novel insight into physiological and biomechanical aspects valuable to coaches, athletes and researchers.
Advanced machine learning technologies are seldom applied to wearable motion sensor data obtained from sport movements. In this work, we therefore investigated neural networks for motion performance evaluation utilizing a set of inertial sensor-based ski jump measurements. A multi-dimensional convolutional network model that related the motion data under aspects of time, placement and sensor type was implemented. Additionally, its applicability as a measure for automatic motion style judging was evaluated. Results indicate that one multi-dimensional convolutional layer is sufficient to recognize relevant performance error representations. Furthermore, comparisons against a Support Vector Machine and a Hidden Markov Model show that the new model out-performs feature-based methods under noisy and biased data environments. Architectures such as the proposed evaluation system can hence become essential for automatic performance analysis and style judging systems in future.
Analysts in professional team sport regularly perform analysis to gain strategic and tactical insights into player and team behavior. Goals of team sport analysis regularly include identification of weaknesses of opposing teams, or assessing performance and improvement potential of a coached team. Current analysis workflows are typically based on the analysis of team videos. Also, analysts can rely on techniques from Information Visualization, to depict e.g., player or ball trajectories. However, video analysis is typically a time-consuming process, where the analyst needs to memorize and annotate scenes. In contrast, visualization typically relies on an abstract data model, often using abstract visual mappings, and is not directly linked to the observed movement context anymore. We propose a visual analytics system that tightly integrates team sport video recordings with abstract visualization of underlying trajectory data. We apply appropriate computer vision techniques to extract trajectory data from video input. Furthermore, we apply advanced trajectory and movement analysis techniques to derive relevant team sport analytic measures for region, event and player analysis in the case of soccer analysis. Our system seamlessly integrates video and visualization modalities, enabling analysts to draw on the advantages of both analysis forms. Several expert studies conducted with team sport analysts indicate the effectiveness of our integrated approach.
Human activity recognition (HAR) aims to recognize activities from a series of observations on the actions of subjects and the environmental conditions. The vision-based HAR research is the basis of many applications including video surveillance, health care, and human-computer interaction (HCI). This review highlights the advances of state-of-the-art activity recognition approaches, especially for the activity representation and classification methods. For the representation methods, we sort out a chronological research trajectory from global representations to local representations, and recent depth-based representations. For the classification methods, we conform to the categorization of template-based methods, discriminative models, and generative models and review several prevalent methods. Next, representative and available datasets are introduced. Aiming to provide an overview of those methods and a convenient way of comparing them, we classify existing literatures with a detailed taxonomy including representation and classification methods, as well as the datasets they used. Finally, we investigate the directions for future research.
Human Activity Recognition has witnessed a significant progress in the last decade. Although a great deal of work in this field goes in recognizing normal human activities, few studies focused on identifying motion in sports. Recognizing human movements in different sports has high impact on understanding the different styles of humans in the play and on improving their performance. As deep learning models proved to have good results in many classification problems, this paper will utilize deep learning to classify cross-country skiing movements, known as gears, collected using a 3D accelerometer. It will also provide a comparison between different deep learning models such as convolutional and recurrent neural networks versus standard multi-layer perceptron. Results show that deep learning is more effective and has the highest classification accuracy.
One of the hot topics in modern era of cricket is to decide whether the bowling action of a bowler is legal or not. Because of the complex bio-mechanical movement of the bowling arm, it is not possible for the on-field umpire to declare a bowling action as legal or illegal. Inertial sensors are currently being used for activity recognition in cricket for the coaching of bowlers and detecting the legality of their moves, since a well trained and legal bowling action is highly significant for the career of a cricket player. After extensive analysis and research, we present a system to detect the legality of the bowling action based on real time multidimensional physiological data obtained from the inertial sensors mounted on the bowlers arm. We propose a method to examine the movement of the bowling arm in the correct rotation order with a precise angle. The system evaluates the bowling action using various action profiles. The action profiles are used so as to simplify the complex bio-mechanical movement of the bowling arm along with minimizing the size of the data provided to the classifier. The events of interest are identified and tagged. Algorithms such as support vector machines, k-nearest neighbor, Naïve Bayes, random forest, and artificial neural network are trained over statistical features extracted from the tagged data. To accomplish the reliability of outcome measures, the technical error of measurement was adopted. The proposed method achieves very high accuracy in the correct classification of bowling action.
The deadlift is a compound full-body exercise that is fundamental in resistance training, rehabilitation programs and powerlifting competitions. Accurate quantification of deadlift biomechanics is important to reduce the risk of injury and ensure training and rehabilitation goals are achieved. This study sought to develop and evaluate deadlift exercise technique classification systems utilising Inertial Measurement Units (IMUs), recording at 51.2 Hz, worn on the lumbar spine, both thighs and both shanks. It also sought to compare classification quality when these IMUs are worn in combination and in isolation. Two datasets of IMU deadlift data were collected. Eighty participants first completed deadlifts with acceptable technique and 5 distinct, deliberately induced deviations from acceptable form. Fifty-five members of this group also completed a fatiguing protocol (3-Repition Maximum test) to enable the collection of natural deadlift deviations. For both datasets, universal and personalised random-forests classifiers were developed and evaluated. Personalised classifiers outperformed universal classifiers in accuracy, sensitivity and specificity in the binary classification of acceptable or aberrant technique and in the multi-label classification of specific deadlift deviations. Whilst recent research has favoured universal classifiers due to the reduced overhead in setting them up for new system users, this work demonstrates that such techniques may not be appropriate for classifying deadlift technique due to the poor accuracy achieved. However, personalised classifiers perform very well in assessing deadlift technique, even when using data derived from a single lumbar-worn IMU to detect specific naturally occurring technique mistakes.
The popularity of running has increased in recent years. A rise in the incidence of running-related overuse musculoskeletal injuries has occurred parallel to this. This study investigates the capability of using data from a single inertial measurement unit (IMU) to differentiate between running form in a non-fatigued and fatigued state. Data was captured from an IMU placed on the lumbar spine, right shank and left shank in 21 recreational runners (10 male, 11 female) during separate 400m running trials. The trials were performed prior to and following a fatiguing protocol. Following stride segmentation, IMU signal features were extracted from the labelled (non-fatigued vs fatigued) sensor data and used to train both a Global and Personalised classifier for each individual IMU location. A single IMU on the Lumbar spine displayed 75% accuracy, 73% sensitivity and 77% specificity when using a Global Classifier. A single IMU on the Right Shank displayed 100% accuracy, 100% sensitivity and 100% specificity when using a Personalised Classifier. These results indicate that a single IMU has the potential to differentiate between non-fatigued and fatigued running states with a high level of accuracy.
To date, the monitoring of fast bowling workloads across training and competition environments has been limited to counting total balls bowled. However, bowling at faster velocities is likely to require greater effort while also placing greater load on the bowler. This study investigated the relationship between prescribed effort and microtechnology outputs in fast bowlers to ascertain whether the technology could provide a more refined measure of workload. Twelve high performing fast bowlers (mean ± SD age; 20.3 ± 2.2 yr) participated in this study. Each bowler bowled 6 balls at prescribed bowling intensities of 60%, 70%, 85% and 100%. The relationship between microtechnology outputs, prescribed intensity and ball velocity were determined using polynomial regression. Very large relationships were observed between prescribed effort and ball velocity for peak PlayerLoad(TM) (R = 0.83 ± 0.19 and 0.82 ± 0.20). The Player Load(TM) across lower ranges of prescribed effort exhibited higher coefficient of variation (CV) [60% = 19.0 (17.0 - 23.0)%] while the CV at higher ranges of prescribed effort was lower [100% = 7.3 (6.4 - 8.5)%]. Routinely used wearable microtechnology devices offer opportunities to examine workload and intensity in cricket fast bowlers outside the normal metrics reported. They offer a useful tool for prescribing and monitoring bowling intensity and workload in elite fast bowlers.
The world of sports intrinsically involves fast and accurate motion that is not only challenging for competitors to master, but can be difficult for coaches and trainers to analyze, and for audiences to follow. The nature of most sports means that monitoring by the use of sensors or other devices fixed to players or equipment is generally not possible. This provides a rich set of opportunities for the application of computer vision techniques to help the competitors, coaches and audience. This paper discusses a selection of current commercial applications that use computer vision for sports analysis, and highlights some of the topics that are currently being addressed in the research community. A summary of on-line datasets to support research in this area is included.
Lunges are a common, compound lower-limb resistance exercise. If completed with aberrant technique, the increased stress on the joints used may increase risk of injury. This study sought to first investigate the ability of inertial measurement units (IMUs), when used in isolation and combination, to (a) classify acceptable and aberrant lunge technique (b) classify exact deviations in lunge technique. We then sought to investigate the most important features and establish the minimum number of top-ranked features and decision trees that are needed to maintain maximal system classification efficacy. Eighty volunteers performed the lunge with acceptable form and 11 deviations. Five IMUs positioned on the lumbar spine, thighs and shanks recorded these movements. Time and frequency domain features were extracted from the IMU data and used to train and test a variety of classifiers. A single IMU system achieved 83% accuracy, 62% sensitivity and 90% specificity in binary classification and a five IMU system achieved 90% accuracy, 80% sensitivity and 92% specificity. A five IMU set-up can also detect specific deviations with 70% accuracy. System efficiency was improved and classification quality was maintained when using only 20% of the top-ranked features for training and testing classifiers.
Deep neural networks (DNNs) are currently widely used for many artificial intelligence (AI) applications including computer vision, speech recognition, and robotics. While DNNs deliver state-of-the-art accuracy on many AI tasks, it comes at the cost of high computational complexity. Accordingly, techniques that enable efficient processing of deep neural network to improve energy-efficiency and throughput without sacrificing performance accuracy or increasing hardware cost are critical to enabling the wide deployment of DNNs in AI systems. This article aims to provide a comprehensive tutorial and survey about the recent advances towards the goal of enabling efficient processing of DNNs. Specifically, it will provide an overview of DNNs, discuss various platforms and architectures that support DNNs, and highlight key trends in recent efficient processing techniques that reduce the computation cost of DNNs either solely via hardware design changes or via joint hardware design and network algorithm changes. It will also summarize various development resources that can enable researchers and practitioners to quickly get started on DNN design, and highlight important benchmarking metrics and design considerations that should be used for evaluating the rapidly growing number of DNN hardware designs, optionally including algorithmic co-design, being proposed in academia and industry. The reader will take away the following concepts from this article: understand the key design considerations for DNNs; be able to evaluate different DNN hardware implementations with benchmarks and comparison metrics; understand trade-offs between various architectures and platforms; be able to evaluate the utility of various DNN design techniques for efficient processing; and understand of recent implementation trends and opportunities.
This paper presents a miniature wearable device and a system for detecting and recording the movement and biometric information of a user during sport activities. The wearable device is designed to be worn on a wrist and can monitor skin temperature and pulse rate. Furthermore it can monitor arm movement and detect gestures using inertial measurement unit. The device can be used for various professional and amateur sport applications and for health monitoring. Because of its small size and minimum weight, it is especially appropriate for swing-based sports like tennis or golf, where any additional weight on the arms would most likely disturb the player and have some influence on the player’s performance. Basic signal processing is performed directly on the wearable device but for more complex signal analysis the data can be uploaded via the Internet to a cloud service, where it can be processed by a dedicated application. The device is powered by a light-weight miniature LiPo battery and has about 6 hours of autonomy at maximum performance.
Many injuries in sports are caused by overuse. These injuries are a major cause for reduced performance of professional and non-professional beach volleyball players. Monitoring of player actions could help identifying and understanding risk factors and prevent such injuries. Currently, time-consuming video examination is the only option for detailed player monitoring in beach volleyball. The lack of a reliable automatic monitoring system impedes investigations about the risk factors of overuse injuries. In this work, we present an unobtrusive automatic monitoring system for beach volleyball based on wearable sensors. We investigate the possibilities of Deep Learning in this context by designing a Deep Convolutional Neural Network for sensor-based activity classification. The performance of this new approach is compared to five common classification algorithms. With our Deep Convolutional Neural Network, we achieve a classification accuracy of 83.2%, thereby outperforming the other classification algorithms by 16.0%. Our results show that detailed player monitoring in beach volleyball using wearable sensors is feasible. The substantial performance margin between established methods and our Deep Neural Network indicates that Deep Learning has the potential to extend the boundaries of sensor-based activity recognition.
Strength and conditioning (S&C) coaches offer expert guidance to help those they work with achieve their personal fitness goals. However, due to cost and availability issues, individuals are often left training without expert supervision. Recent developments in inertial measurement units (IMUs) and mobile computing platforms have allowed for the possibility of unobtrusive motion tracking systems and the provision of real time individualised feedback regarding exercise performance. These systems could enable S&C coaches to remotely monitor sessions and help gym users record workouts. One component of these IMU systems is the ability to identify the exercises completed. In this study, IMUs were positioned on the lumbar spine, thighs and shanks on 82 healthy participants. Participants completed 10 repetitions of the squat, lunge, single leg squat, deadlift and tuck jump with acceptable form. Descriptive features were extracted from the IMU signals for each repetition of each exercise and these were used to train an exercise classifier. The exercises were detected with 99% accuracy when using signals from all five IMUs, 98% when using signals from the thigh and lumbar IMUs and 98% with just a single IMU on the shank. These results indicate that a single IMU can accurately distinguish between five common multi-joint exercises.
Objectives:
Collision frequency during rugby league matches is associated with team success, greater and longer lasting fatigue and increased injury risk. This study researched the sensitivity and specificity of microtechnology to count collision events during rugby league matches.
Design:
Diagnostic accuracy study.
Methods:
While wearing a microtechnology device (Catapult, S5), eight professional rugby league players were subjected to a total of 380 collision events during matches. Video footage of each match was synchronised with microtechnology data. The occurrence of each collision event was coded in comparison with whether that event was or was not detected by microtechnology.
Results:
Microtechnology detected 371 true-positive collision events (sensitivity=97.6±1.5%). When low-intensity (<1 PlayerLoad AU), short duration (<1s) events were excluded from the analysis, specificity was 91.7±2.5%, accuracy was 92.7±1.3%, positive likelihood ratio was 11.4×/÷1.4 and the typical error of estimate was 7.8%×/÷1.9 (d=0.29×/÷1.9, small). Microtechnology collisions were strongly and positively correlated with video coded collision events (r=0.96). The ability of microtechnology to detect collision events improved as the intensity and duration of the collision increased.
Conclusions:
Microtechnology can identify 97.6% of collision events during rugby league match-play. The typical error associated with measuring contact events can be reduced to 7.8%, with accuracy (92.7%) and specificity (91.7%) improving, when low-intensity (<1 PlayerLoad AU) and short duration (<1s) collision reports are excluded. This provides practitioners with a measurement of contact workload during professional rugby league matches.
The increasing popularity of wearable devices in recent years means that a diverse range of physiological and functional data can now be captured continuously for applications in sports, wellbeing, and healthcare. This wealth of information requires efficient methods of classification and analysis where deep learning is a promising technique for large-scale data analytics. Whilst deep learning has been successful in implementations that utilize high performance computing platforms, its use on low-power wearable devices is limited by resource constraints. In this paper, we propose a deep learning methodology, which combines features learnt from inertial sensor data together with complementary information from a set of shallow features to enable accurate and real-time activity classification. The design of this combined method aims to overcome some of the limitations present in a typical deep learning framework where on-node computation is required. To optimize the proposed method for real-time on-node computation, spectral domain pre-processing is used before the data is passed onto the deep learning framework. The classification accuracy of our proposed deep learning approach is evaluated against state-of-the-art methods using both laboratory and real world activity datasets. Our results show the validity of the approach on different human activity datasets, outperforming other methods, including the two methods used within our combined pipeline. We also demonstrate that the computation times for the proposed method are consistent with the constraints of real-time on-node processing on smartphones and a wearable sensor platform.
As is well known, the target recognition algorithm of hybrid information system has intrinsic disadvantages, such as high time complexity, high performance requirements of hardware and complex operations, in this paper, a fast golf gesture recognition algorithm of static image and video sequence is proposed for the field of sports auxiliary training. In static image recognition, a fast multi-scale aggregation channel feature is utilized to extract hybrid information, and the extraction speed can be improved through an approximate calculation method. An improved AdaBoost classifier is adopted to classify the information. On this basis, the aggregation of channel feature detector locates the prominence region of static image, and then scans the generated fractional sequence through the gesture detector as the feature data of golf gesture in the video sequence. Finally, the real-time judgment of feature data is carried out with a linear support vector machine, the rapid identification of golf swing gesture can therefore be obtained. The experimental results show that the recognition speed is over 30 fps and the accuracy is 97% on iPhone5s and later versions, which suggest the validity of algorithm in practical application.
Inertial sensors are powerful motion measurement devices. They are well-known in vehicle guidance and enable a detailed capture of position, attitude, velocity, and acceleration. Due to modern technology, navigation systems based on these sensors became increasingly small, light, and inexpensive. So, they suggest themselves for motion analysis in sports as an arising application area. Considering the last decades, this paper outlines and discusses the introduction and typical usage of inertial and integrated navigation systems in sports and biomechanics respectively. --- Springer Nature Content Sharing Link: https://rdcu.be/bpGPk
The world of sports is being transformed by the rise of automated systems which enable the monitoring of athletes. Such systems provide objective assessments about the performance of athletes, the loads they are exposed to or their tactical behavior. P