Improved Pedestrian Detection Algorithm of Yolov4 Network Structure

When the YOLOV4 network detects pedestrians alone, the small target pedestrians will be missed, resulting in the reduction of P (Precision) and AP (Average Precision) values. This paper improves the YOLOV4 network structure. In order to improve the feature extraction capability of the network for small targets, a shallower feature layer is added to the original three output feature layers of the YOLOV4 backbone network to build PANet (Path Aggregation Network) together. And two SPP (Spatial Pyramid Pooling) structures are added to expand the receptive field. The channel attention mechanism module is added and some convolutional layers of the original network are deleted. Finally, transfer learning is used to make the detection effect better. The P value of the pedestrian on the PASCAL VOC data set increased from 84.43% to 91.37%, and the AP value increased from 74.78% to 87.39%, and the P value on the commonly used pedestrian detection data set INRIA (INRIA Person Dataset) increased from 93.20% increased to 98.02%, AP value increased from 91.08% to 94.02%. Experimental results show that the network has a better effect on pedestrian detection, and the accuracy and average precision are improved.