Research on Lightweight Pedestrian Detection Model in Complex Background

Abstract

Pedestrian detection under complex background is a hot research content in the field of Computer Vision. Pedestrian detection in a real and complex environment is extremely important for Human Action Recognition, Intelligent Transportation, Video Surveillance Security and other security aspects. This paper proposes a lightweight model based on the YOLO algorithm for pedestrian detection in a complex background. First, we build a YOLO v3 model based on Darknet-53. Then in order to simplify the pedestrian detection model and ensure the performance of the model, this paper adopts the method of convolution channel pruning to lighten the model. This method realizes rapid and accurate detection of pedestrians in complex backgrounds. In order to verify the performance of the proposed model, this paper constructs a scientific and reasonable small dataset sample based on the YOLO model. The dataset contains 1505 images, covering various complex scenes such as intersections with crowded people and traffic at night, and commercial streets with complex buildings. The test results show that for a complex background dataset with a small number of samples, the model's mAP reached 87.89%, precision was 90.37%, and recall was 77.05%. In addition, this paper also compares the performance of YOLO v3 and YOLO v4-tiny models. The results show that the lightweight models proposed in this paper have strong robustness. It is feasible to apply this method to real-time and accurate detection of pedestrians in complex backgrounds.