Research On Occlusion Pedestrian Detection Algorithm Based On Deep Learning

As for the pedestrian detection task under occlusion,the detector’s performance can be severely degraded due to varying degrees of occlusion on the targets.To address the detection performance issues caused by missing target features under occlusion,it is necessary to enhance the visible feature information,reduce the impact of occlusion information on detection,enhance the detection model’s robustness to different occlusion conditions,and thereby improve the accuracy of pedestrian detection while reducing the false negative rate.In this paper,we investigate and analyze the implementation methods of the feature pyramid network used for feature fusion and the backbone network used for feature extraction,and propose improvements to these networks to tackle the occlusion problem.Based on the Faster R-CNN（Region-Convolutional Neural Networks）detection algorithm,this paper designs a new module to enhance the detection performance of occluded pedestrians.The main innovations are as follows:Firstly,in response to the problem of reduced pedestrian feature information under occlusion conditions,this paper proposes a visible region enhancement network module based on the attention mechanism.This module distinguishes the information in the feature map into visible feature information and interference feature information,adjusts the feature information through the attention mechanism,enhances the visible information,suppresses the interference information,and finally outputs the adjusted feature map to the subsequent detection module for final classification and regression positioning.Secondly,in response to the problem of multi-scale feature information variation under occlusion conditions,the feature pyramid network can effectively improve the feature representation ability and achieve better detection accuracy.However,feature extraction for occluded targets remains a limitation,and the increasing depth of the pyramid network and complex fusion methods increase the number of network parameters,which reduces the efficiency of the network.Moreover,the impact of feature information from different resolution feature maps on the output features is usually not equal.To address these limitations,this paper proposes a weighted feature pyramid network,which adds a weight network to the feature pyramid network.When fusing two different hierarchical features,the weight network determines their weight proportions based on the importance of each feature layer,and then the weighted feature layers are fused.Finally,based on the first two works,this paper designs an improved Faster R-CNN algorithm by combining the weighted feature pyramid network and visible region enhancement network.The improved algorithm combines the advantages of both modules and can better cope with occlusion scenes.Experimental results on the Crowd Human pedestrian dataset show that the two proposed modules can effectively improve the detection performance of the algorithm under occlusion conditions compared to existing detection algorithms.Compared with the benchmark network,the combined algorithm improves the AP,MR^-2 and JI indexes by 2.45%,2.11%and 1.22%,respectively.In summary,the algorithm proposed in this paper provides an effective direction for future research.