Improved Pedestrian Detection Algorithm Based on YOLOv5s

Abstract

In this study, we propose YOLOv5s-PGD algorithm for dense pedestrian detection, which can improve the recall and reduce the number of parameters compared with YOLOv5s. First, a minimum scale detection layer has been added to deepen the pyramid’s depth and enhance detection accuracy. Second, ghost convolution has been employed to replace standard convolution to increase real-time performance of the algorithm. Finally, depth separable convolution has been used to address issues of high parameters and large computational complexity that lower the efficiency of the algorithm. Experiment results demonstrate that the detection accuracy of the YOLOv5s-PGD algorithm on the CrowdHuman public dataset is up to 85.1%, which is 2.2% higher than that of YOLOv5s. Furthermore, the number of parameters has decreased by 19.7%, and the calculation burden has decreased by 2.5%. Consequently, the proposed YOLOv5s-PGD algorithm better satisfies the requirements of real-time detection, model optimization, and terminal deployment in dense pedestrian scenarios.