A pedestrian detection algorithm based on improved YOLOv2
Ziwei Liu, Ying Shi, Mingjun Sun · 2018
For improving the accuracy of pedestrian detection, an improved algorithm based on YOLOv2 network framework is proposed. Usually, a large number of redundant candidate proposal boxes of detected pedestrians exist via using YOLOv2 framework. In this paper, clustering algorithm is adapted to obtain the priori knowledge of the pedestrian target proposal box scale and aspect ratio in order to formulate the proposal box filtering rules and to remove the redundant filtered proposal boxes. Experimental results show that on the KITTI dataset, the performance of improved algorithm will degrade with the increase of difficulty level of targets, but the precision and recall rate increases obviously and the detection accuracy increases respectively by 9.03%, 6.37% and 5.91% for the easily, moderately and hardly detected targets compared to YOLOv2.