Pedestrian Detection Based on Modified YOLOv5

Abstract

In the pedestrian detection scenario, the detection algorithm usually misses obscured and distant fuzzy pedestrians, and at the same time cannot take into account the detection accuracy and speed. In this paper, we propose a modified YOLOv5 model for pedestrian detection. Firstly, the backbone network uses the SPD-GCONV module constructed by the combination of SPD (Space-to-Depth) module and Ghost convolution for down-sampling to reduce the loss of fine-grained feature information. Secondly, the multi-scale detection ability of the model is enhanced by adding a small size detection layer. Then, the original CIoU loss function is replaced by $\alpha$-EloU loss function to improve the accuracy of pedestrian target location. According to the experiments on WiderPerson data set, the average detection accuracy is improved by 2\% compared with other pedestrian detection algorithms on the premise of ensuring the detection speed. Experimental results show that the improved algorithm can significantly improve the detection performance.