Track Obstacle Detection Algorithm Based on YOLOv3
Zijian Cong, Xiaoguang Li · 2020
In this paper, YOLOv3 algorithm is applied to track obstacle detection to achieve the robustness of detection. Firstly, the network structure and principle of YOLOv3 algorithm are introduced. Then, the boundary box prediction function and category prediction of multi-scale target detection of YOLOv3 are analyzed. Finally, the implementation steps of YOLOv3 are given. Experimental results show that, compared with YOLOv2 algorithm, YOLOv3 algorithm has better robustness, more targets can be detected and positioning is more accurate in the case of complex background with good lighting and low illumination contrast. In the case of more pedestrians in front of the track in a good light environment, the number of detected targets increased from 4 to 8, and the average prediction probability of the target category increased from 0.673 to 0.978. For the complex background with human-vehicle mixture, the average prediction probability increased from 0.72 to 0.94. In the case of poor illumination environment in rainy dusk, the detection target increased from 2 cars to 4 cars and 1 pedestrian, and the average forecast probability increased from 0.51 to 0.77. For the small target image of two pedestrians in the distance in front of the track, YOLOv2 algorithm cannot detect the target, while YOLOv3 algorithm can accurately detect the target.