Research on Lightweight Pedestrian Detection Method Based on YOLO

Abstract

Aiming at the problems of large size, high calculation cost and slow detection speed of current pedestrian detection models, this paper proposes a lightweight improved pedestrian detection algorithm based on YOLO v5. Firstly, the Shufflenet v2 network is introduced to replace the backbone network of YOLO v5. Then cascade convolution is designed, and the size of the backbone extraction network convolution core is modified to improve the sensing field of the backbone feature extraction network so that more important context features can be separated. Finally, the unnecessary structure of the backbone network is cut to reduce the scale of network parameters and improve the inference speed. In this paper, the INRIA dataset is used for relevant experiments. Through the experimental analysis of the two algorithms, the size of the model, the number of parameters and the reasoning time of the algorithm in this paper are reduced to 50.1%, 48.6% and 64.7% of YOLO v5s model, respectively. In contrast, the average accuracy of the algorithm is only reduced by 2.1%. This algorithm not only guarantees accuracy, but also greatly improves the reasoning speed.