Research on Algorithm of Pedestrian Detection in Dense Crowd Scene

• 1、School of Automation,Beijing University of Posts and Telecommunications,Beijing,100876

Abstract：Pedestrian detection has a very wide range of applications in computer vision. It has become a hot topic that has attracted much attention. However, due to the differences between human appearance and serious occlusion of pedestrians, the traditional image processing methods cannot meet the needs of practical applications.This paper focuses on computer vision technology in artificial intelligence. It proposes object detection algorithms based on deep learning and applies them to dense crowd scenes to complete pedestrian localization tasks. This paper improves the single-stage detector RetinaNet and uses the training method of transfer learning to achieve fast and accurate recognition of pedestrians. The final average recognition accuracy rate is more than 95%, with mAP reaching 72.16%. And the recognition time of a single image only needs 0.04 seconds. Then, this paper proposes a loss function with a penalty term, which can effectively solve the problem of pedestrian occlusion commonly encountered in urban scenes, Also it further improves the pedestrian detection mAP to 73.45%, while the precision is 95.03%, and the recall is 86.37%，which can meet the need of application on pedestrian.

Keywords： Neural Networks Loss Function Pedestrian Detection