Pedestrian Detection And Behavior Discrimination Based On Convolutional Neural Network

With the arrival of the information age in China,pedestrian detection,behavior discrimination and other technologies have attracted the attention of scholars.Although there are mature schemes for pedestrian detection and behavior discrimination in the field of computer vision,there are still unsolved problems,such as pedestrian targets are vulnerable to various effects such as shooting angle,shooting distance,occlusion and target movement,resulting in small targets to be measured,irregular target shape and similar actions.In view of the above problems,this paper studies the pedestrian objectives in the urban rail transit scene,uses the method based on deep learning to realize the detection and behavior analysis of pedestrian head and shoulder,and optimizes the network structure to improve the accuracy of detection.The main work and innovation of this paper are as follows:(1)In order to solve the problem of missing detection and misdetection caused by serious occlusion in target detection,YOLOv5 s target detection algorithm is selected as the research framework of this paper and improved.First of all,the data set of pedestrian head and shoulder as the target is constructed.Secondly,the feature fusion module in the network is analyzed,and the underlying information of different network layers is fully utilized and fused.Then,a deformable convolutional network is added to the feature extraction network to solve the problem of irregular shape caused by target occlusion.At the same time,the feature extraction performance of the network is improved by introducing the attention mechanism module and giving different weights to the feature information.Finally,verification is carried out on the self-made pedestrian head and shoulder data set.Experimental results show that in the subway station scenario,the improved network improves the detection accuracy of small pedestrian head and shoulder targets,with m AP increasing from 91.86% to94.12%,an overall improvement of 2.26%.(2)In terms of behavior discrimination,conduct behavior recognition research on the head and shoulder area after target detection,distinguish smoking,telephone and other behaviors.Due to the low pixels of images in the data set,the recognized behaviors have problems such as fuzzy and similar actions.In this paper,Res Net network is selected as the behavior classification network in this paper.Starting from the Loss function,label smoothing,Focal Loss and Center Loss functions are introduced to improve the classification accuracy.The block convolution module is adopted to reduce the amount of network computation.Through experimental comparison,both loss functions can improve the performance of the network,and the effect is the best when acting on the network at the same time,and the accuracy rate is improved by 1.29 percentage points,reaching 96.91%.Secondly,the behavior recognition system and interface are designed based on the improved algorithm.The experimental test shows that the system can meet the real-time requirements.Through the optimization of the network and the analysis of the experimental results,the method proposed in this paper has been improved in terms of detection accuracy and other performance,and has good applicability in practical scenarios.