Abstract: When pedestrian detection is carried out in image and video sequences, there are some problems, such as diverse pedestrian posture and scale as well as pedestrian occlusion, which leads to inaccurate detection of some pedestrians by YOLOv4 algorithm and false detection and missed detection. In order to solve this problem, a complex pedestrian detection model based on improved YOLOv4 algorithm is proposed. Firstly, the ground truth size of pedestrian dataset is analyzed by the improved k-means clustering algorithm, and the size of anchor box is determined according to the clustering results. Secondly, PANet is used for multi-scale feature fusion to make it more sensitive to multi-attitude and multi-scale pedestrian targets and improve the detection effect. Finally, for pedestrian occlusion problem, the repulsion loss function is proposed to make the predicted box as close to the correct target as possible. The experimental results show that the new de-tection model has better detection effect than YOLOv4 and other pedestrian detection model.

Key words: pedestrian detection, YOLOv4 algorithm, k-means clustering algorithm, multi-scale feature fusion, repulsion loss