Activity Recognition - Science topic

means the design of the model, and how you train and validate your model, and then test it. it also includes the data preprocessing steps and feature engineering.

Dear Shanaka Ramesh Gunasekara ,

data set i found :

WISDM

PAMAP2

UniMib shar

usc had

You may encounter over-fitting which may impede accurate identification of unseen features.

Donald E. Watenpaugh Such a nice concern about working communities or individuals. Everybody have information about obesity but sitting for long can lead to muscle atrophy that observed in long bed rest patients or the astronauts. While sitting, anti-gravity bones and muscles relaxed and gravity influences decreases, similar phenomenon observed in space. As per the literature, bad lifestyle decreases bone density by 10% per 3-6 months. 6-9 hour sitting is playing havoc with workers life. Govt or company should provide compulsory walk or stretching sessions.

I had to perform segmentation of human speech for the purpose of my PhD. In my experience, you're better off to keep your original signal/data, and segment it prior to other processing : your decision will be cleaner, and reversible at the same time.

My own real time segmentation was based on the detection of biological events (glottal closure) comparing two sliding models, one short, one long, and using a cumulated sum of Kullback-Leibler divergence estimates, looking for convexity jumps in this criterion (D2f(t) = f(t+1) - 2f(t) + f(t-1) estimates the second order derivative of signal f...).

Get to know the semantics of what you are looking for, the the models will fall in place smoothly

Hope this helps

Ref: my IEEE Journal article on real time speech segmentation and coding, posted on researchgate

Dear Imran Jamal,

I suggest you to see links and attached files on topic.

Best regards

Dear Muhammad Faidhan Elahi ,

Look the link, maybe helpful.

Regards,

Shafagat

See: Johnson, C and Johnson, R. (2000) Alzheimer's disease as a trip back in time. American Journal of Alzheimer's Disease April or the short British Alzheimer's Society one page article. for an explanation of the time Travel model of AD. The updated version with more positive language in in the 2017 Behavioral Science journal article.

Dear Nafees Ahmad ,

Look the link, maybe useful.

Regards,

Shafagat

you can read this research I hope it will be useful for you and good luck

After the long research and discussion, I have found this:

Some loss of data can occur, based on weighted average dynamics - if the weighted average points are not of a representative quantity.

Whilst the average of a entirely symmetrical and perfectly balanced set, indeed, would be partitioned correctly - You’re going to incur some small loss of information - depending on how spread out the points are.

You can counter-act that, with just increasing the interpolation points - But at that point - you’re incurring more costs of processing.

So, i guess the “side effect”, would mostly be a function of Interpolation points contra Accuracy contra parsing speed.

Another point of contention - that could arise - is representation of Co-ordinate points based on Dimensions.

Lower dimension points may inaccurately portray the dimensionality of the problematique - in so far as not being entirely accurate.

But in so far, as inferrence analytics goes and finding the average - sullying out averages from arbitrated interpolation points - enacts as a information strengthener - not a dampener.

Because it fullfills the process of Fisher Metrics of Aggregation - shuffling and re-sampling and running the process.

This is basically how most aggregate Simulation processes are meant to play out - like Monte Carlo and others.

You’re basically meant to aggregate to infinity and “parse in information as you go along” - so, shuffling and updating/resampling - just gives you more accuracy in so far as utilizing Bootstrapping (averaging runs).

So - i guess, the side-effect, would be increased accuracy of Sampling - as you’re better fullfilling the conditional of truly arbitrary sampling (in so far as to account for more information of the collective Distribution of the entire Metric space being analyzed).

But - other than that, you’re not really “losing” anything - in that sense.

Just adding to the process with more power and more data - takes more time.

At least, this is to my understanding of the problematique.

I like your studies.

In my Thar desert, Pakistani culture Camel milk a treatment for ASD (Autistic Spectrum Disorder). Always hope for the best.

Mujeeb

I found the Sussex-Huawei Locomotion Dataset

For Scaling I would say,

Pros:

-Better performance i algorithms like SVMs and Neural networks

-faster conversions

Cons

-you may forget to look for the true coef. in regressions

-won't always guarantee better results

CNN

Respected Mr. Abdessamad Tafraouti,

I use your suggested code but unfortunately It cannot work. Actually text files ( skeleton_pos.txt, a01_s01_e01_realworld & a01_s01_e01_screen ) contains joints locations but i am unable to use them.

Yes Doug P L Hunt it's exactly what I have done. And the time of the day feature improved the results of 20%!

I suspect that your question is too general. My wrist movement is somewhat independent of my body movement. For example, I may be playing a guitar and so my wrist will be moving in a certain manner. However, while playing the guitar I may be either, sitting down, standing up, walking or dancing to the music that I am playing. From this you can see that my wrist movements do not define my body movements and so one can not be inferred from the other.

Another situation where this "confusion" occurs can be manufactured if you have a chair that rotates. You can then sit in the chair with your wrist still while a friend quickly rotates the chair and (if you don't get dizzy) then if you look at the reading from your wrist sensor you will see that it seems to indicate that the wrist is moving (and it is but so is your whole body) and so you can not assume that a movement registered at the wrist only comes from the wrist.

Insert records of activities with start time and NULL value for end time

Dear Ali,

Please go through some of the referred papers given below, which might help you in your research.

1. Adiga, M. T., & Bhandarkar, R. (2016, October). Improving single frequency filtering based Voice Activity Detection (VAD) using spectral subtraction based noise cancellation. In Signal Processing, Communication, Power and Embedded System (SCOPES), 2016 International Conference on (pp. 18-23). IEEE.

2. Sohn, J., Kim, N. S., & Sung, W. (1999). A statistical model-based voice activity detection. IEEE signal processing letters, 6(1), 1-3.

3. Sohn, J., & Sung, W. (1998, May). A voice activity detector employing soft decision based noise spectrum adaptation. In Acoustics, Speech and Signal Processing, 1998. Proceedings of the 1998 IEEE International Conference on (Vol. 1, pp. 365-368). IEEE.

Thanks,

Sobhan

Dear Hasan Khan Lodhi,

Look the link, may be useful.

Regards, Shafagat

I didn't manage to figure it out and dropped the issue. The admin of the dataset did answer eventually, but his answer didn't really help me make sense of it, and I had already stopped using it. I suggest trying to get in touch with him, perhaps he can help you.

thank you Samer for clarification. Now my confusion is resolved

hi

i think gesture recognition is focus on special part of body like hand or head .

in gesture recognition, we dont need to follow all of the body parts , we know gestures

and know that which part of body contains in this gesture .

but in activity detection, it is posible someone might do it differently than I do.

in fact when to use gestures to command , should do it right and i or you should do it almost the same . but in activity we do not have to do alike.

I plan to analyze very specific video segments from very specific movies that some are available online, while others are not. Some of the movies that are on my list are: Metropolis (1927), Blade Runner (1982), Brazil (1985) and The Grand Budapest Hotel (2014).

These methods are somehow old. Now linear mixed effects models offer a wider range of error/random effect correlation patterns and fits data very well. I suggest it.

Could you be a bit more specific on the activities and how the measurements relate to the stress recognition ?

I ask this, because it is a bit complicated to answer your question without some further insight.

kind regards,

Marian

Babak,

This is definitely an interesting question.  I am also unsure of what exactly the goal of your analysis would be, however, if you are interested in getting a sense for a prototype movement, rather than actual live movements, then I would suggest perhaps thinking about it from more of an artist/animator perspective.  What would be the minimum amount you would need to "draw" in order to show climbing?  or running?  or walking?  I think a animated 2D stick figure could easily represent all of these quite well.  This would make your task much easier in terms of the feature set size (degrees of freedom) and the computational complexity.

It would probably be computationally tractable to run a massive PCA/SVM/clustering operation on a huge set of animated 2D stick figures.  Of course, you would have to create the figures somehow...

Hope that is helpful.

Sean

Hi Mehdi. I was the chair of the ECML 2015 Discovery Challenge. We published the dataset of the competition in the UCI repository. It has gps traces of the taxi busy services for a period of one year. It has no labelled visited places...but you can easily label some of them (such as downtown, train station or airport) with a simple haversine distance function. The granulairty of the dataset is 15 seconds.

I hope that this helps.

Best,

Luis

Probably it also depends on the sensing interfase you going to use. For instance, you can partitioning a video area into a grid, then using Markov random fields or some spectral method for graphs. I hope this help...

HMM's are meant exclusively where sequence of events or states define your problem. hence represent each activity as a state and sequence of these states into activity. This helps

Quality function deployment (QFD) is a structured method to extract quantitative parameters from qualitative user requirements. It helps to prioritise the parameters to produce justified decisions. In this case, the network selection.

Whereas, QoS is a quantitatively measured performance of network. Specifically, the performance perceived by the users of the network.

Just a suggestion to try Sir , since is not my area :

- LVQ

- Intrinsic distance

- Motion History Gradients

Good luck sir :)

Hello. I imagine that it depends what you mean by abnormal event. From my point of view, this type of problem takes two forms. The first is an image processing problem where you look in a single frame for an object with a certain feature. This can be reduced to a parameter estimation problem. The second way is to view the problem as one of change detection, i.e., the abnormal event corresponds to a change in the dynamics of the underlying system (think of you car engine when something inside it breaks). There is a large literature on change detection, which can be approached in many ways. My experience of this type of problem is from the area of manoeuvring target tracking. Hopefully you can find some useful references for your work as quite a lot has been done in this area (see link), although the sensor is usually a radar system rather than a video sequence.

In Judaism there is often, especially in Orthodox weddings, a document that is in addition to the Ketubah, the marriage contract.  This document is called Tanayim, conditions, and it often contains religious aspects, such as the foods in the home, religious observance, training of children, etc. 

I use a biotin/ antibody ratio 4/1 (Mol to mol). Using this ratio is very unlikely that you see any alteration in antigen recognition. A ratio higher tan that (i.e. 6/1) frequently gives problems

this one is also related to your quastion

Do you want to classify the activity (running, jumping, sleeping) or just to detect any movement?

Dear Hussein,

There is no classifier which is always the best.

In the very cited paper

[

Poppe, R. (2010). A survey on vision-based human action recognition. Image and vision computing, 28(6), 976-990.‏

It can be uploaded from

]

you can find there (in chapter 3) a summary of many machine learning methods that are used for human action recognition.

You can see there methods form different categories, e.g.:

Discriminative classifiers such as SVM, RVM and AdaBoost

Nearest neighbor methods,

Discriminative models such as HMM and CRF

and many other kinds of ML methods.

There are also specific papers that describe a use of a certain ML method, such as:

HMM:

Yamato, J., Ohya, J., & Ishii, K. (1992, June). Recognizing human action in time-sequential images using hidden markov model. In Computer Vision and Pattern Recognition, 1992. Proceedings CVPR'92., 1992 IEEE Computer Society Conference on (pp. 379-385). IEEE.‏

SVM:

Schuldt, C., Laptev, I., & Caputo, B. (2004, August). Recognizing human actions: a local SVM approach. In Pattern Recognition, 2004. ICPR 2004. Proceedings of the 17th International Conference on (Vol. 3, pp. 32-36). IEEE.‏

Yaakov

Rajeev.

TPR and FPR separately do not tell you anything. The same goes for Precision and Recall. F-score, Equal Error Rate and similar can give you some information, but it is limited and you have to be careful how you use those values.

Look at this previous question:

I'll copy part of my answer for that question here:

You have to first answer a simple question: Do You know what are the costs of the decisions?

If the costs are known, you can calibrate your classifiers (set the operating point), and compare costs of the classifiers.

If you do not know the costs, you may have some insight about a reasonable range of operating points of the target application. Lets say the target application is a surveillance system, and all positive responses will have to be shown to a human operator which is able to process one event per minute. In such case, it makes sense to compare detection rates at one false positive per minute. You can use other measures instead, but it makes sense to keep one type of error fixed and compare the other one.

If you have no knowledge about the possible operating point in the target application, you should show results in full range of possible operating points. Plot ROC, precision-recall, DET or something similar. You can also compute area under ROC or Precision-recall and get a numeric value which reflects performance over the whole range of regimes; however, such aggregation has its own problems.

Seems like an interesting problem, but I don't think that a deep neural net is what you really want to use. However, here are some bibliography links I have on the subject of Spatial-Temporal analysis for Gesture Interpretations:

You can also use some of these datasets to train:

Another thing you might want to look into is the area of Content-Based Video Retrieval, that deals with the same problem of reducing image sets into features that could be indexed and later be retrieved. While not a lot of cohesive literature around, here is a paper I like to use as introduction:

Hopefully this helps!

Thank you for the links. How about using grab-cut for detecting humans and graph-cut method for detecting and extraction of sihouettes

I have not worked with Microsoft SDK for face tracking but the SDK available for skelton tracking, Depth tracking they are quite robust in nature which I am using it in my project

OK I will look at it

I dont think that skin detection going to play a major role in human action recognition..... because skin detection can only be used like in face detection like or face tracking options.... but it cant be efficiently used in action recognition......... If u r using skin detection for FG/BG extraction then its fine, but u will only extract the human being as a FG but no action recognition using skin detection in my opinion...

Konstantinos,

I agree that most work on human activity has assumed pre-segmented data and addressed only the recognition part of the task. There are exceptions, however. To toot our own horn, at Colorado State University (as part of DARPAs Mind's Eye program) we tackle both detection and recognition together and as an unsupervised learning problem. It has been hard so far to get the whole system published (but see O'Hara & Draper CVPR 2012 & O'Hara & Draper WACV 2013 for pieces). The system is keyed by motion: moving objects are tracked, and a sliding window cuts video snippets from the tracks of moving objects. At training time, the snippets are clustered to form "activities". At performance time, activities in new tracks are labeled according to these activities. The goal is take a pool of unlabeled videos, learn what activities occur multiple times, and then be able to spot those activities in novel (unlabeled) videos.

Not yet. But currently we are working on acquiring an official permit.

You can also check our project web site :