Hanmin Lee’s research while affiliated with Korea Institute of Machinery and Materials and other places

Tire-road friction coefficient information is an essential factor in the driving stability and safety of a vehicle. In recent years, there has been a lot of research on using the vibration characteristic of tires to estimate the road surface condition from its features. However, since tire vibration characteristics vary depending on conditions such as tire pressure, load, and driving status, it is still difficult to develop a road surface classification algorithm that is robust to various situations. To overcome this limitation, this paper proposes a road surface classification algorithm using a one-dimensional convolutional neural network (CNN) based on acceleration signals obtained through an intelligent tire sensor attached inside the tire. Moreover, a time series data augmentation method is applied to ensure that the learning network has the robustness to perform well under different tires and driving conditions than that in the training dataset. A road surface classification algorithm is trained using a dataset of accelerations measured on dry asphalt, wet asphalt, and basalt tile roads, and the performance of the trained algorithm is validated through test scenarios considering different tire conditions and vehicle types. Furthermore, the performance of different CNN architectures is compared and the algorithm with the best performance is suggested. The robustness to different tires and driving conditions makes the proposed algorithm practical for estimating road surface conditions in real vehicles.

As unmanned ground vehicles (UGV) continue to be adapted to new applications, an emerging area lacks proper guidance for global route optimization methodology. This area is surveillance. In autonomous surveillance applications, a UGV is equipped with a sensor that receives data within a specific range from the vehicle while it traverses the environment. In this paper, the ant colony optimization (ACO) algorithm was adapted to the UGV surveillance problem to solve for optimal paths within sub-areas. To do so, the problem was modeled as the covering salesman problem (CSP). This is one of the first applications using ACO to solve the CSP. Then, a genetic algorithm (GA) was used to schedule a fleet of UGVs to scan several sub-areas such that the total distance is minimized. Initially, these paths are infeasible because of the sharp turning angles. Thus, they are improved using two methods of path refinement (namely, the corner-cutting and Reeds–Shepp methods) such that the kinematic constraints of the vehicles are met. Several test case scenarios were developed for Goheung, South Korea, to validate the proposed methodology. The promising results presented in this article highlight the effectiveness of the proposed methodology for UGV surveillance applications.

Dual crane lifting, wherein two cranes collaborate to lift a single workpiece, serves as an essential solution in scenarios in which employing a single, sufficiently large crane is impractical due to cost constraints, ground conditions, and spatial limitations. Due to the complexity of double crane lifting operations, the implementation of automated path generation minimizes the risk of human error and removes the potential for accidents by simulating and validating the generated crane path. We propose a novel multiple rapidly-exploring random trees (RRT) based algorithm designed specifically for dual crane systems to produce lifting paths, particularly in challenging ’narrow path finding’ scenarios. The multiple RRT method is an efficient way to find paths in environments with high complexity and low connectivity through a strategy that allows new trees to be generated and grown whenever a newly generated node that cannot be connected to an existing tree occurs. The proposed path planning algorithm not only adapts the multiple RRT method to the dual crane systems but also incorporates ideas to enhance the optimality of generated paths while reducing computational time. The effectiveness of this algorithm has been validated through a case studies covering various scenario.

Light detection and ranging (LiDAR) sensors can create high-quality scans of an environment. However, LiDAR point clouds are affected by harsh weather conditions since airborne particles are easily detected. In literature, conventional filtering and artificial intelligence (AI) filtering methods have been used to detect, and remove, airborne particles. In this paper, a convolutional neural network (CNN) model was used to classify airborne dust particles through a voxel-based approach. The CNN model was compared to several conventional filtering methods, where the results show that the CNN filter can achieve up to 5.39 % F1 score improvement when compared to the best conventional filter. All the filtering methods were tested in dynamic environments where the sensor was attached to a mobile platform, the environment had several moving obstacles, and there were multiple dust cloud sources.

The paper presents a novel method for real-time detection of off-road drivable regions in LiDAR point clouds without relying on deep learning. Our method directly distinguishes drivable regions while accounting for noise caused by snow and fog. By focusing on the scanning characteristics of rotating LiDAR, our PIIC approach achieves robust performance in challenging off-road environments. The proposed technique divides the space into sections and segments drivable areas by counting points in each section. It allows noise influence to be countable using a simple probabilistic model. The developed solution offers significant advancements in autonomous driving in off-road environments since the method offers drivability in real-time with noise immunity. Experimental results demonstrate the efficiency of the proposed method in hazardous situations such as heavy snow, fog, and crosstalk from other LiDAR signals. Importantly, the low computational load enables real-time operation and provides promising opportunities for fusion with other recognition methods.