Research On Pedestrian Feature Analysis And Re-identification Methods Based On Deep Learning

Person re-identification,a technique that uses computer vision techniques to determine whether people in different video sequences or images are the same target,is a hot-button issue in the field of computer vision that has both research value and challenges.Due to the interference of people being blocked,distorted,blurred,etc.in the video image,person re-identification algorithm has low feature utilization rate,low recognition accuracy and other problems,which is difficult to meet the needs of practical applications.Currently,many approaches based deep learning can effectively improve the accuracy of person re-identification and have received extensive attention from many scholars.This paper focuses on the person re-identification method based on deep learning,which conducts research on pedestrian detection,feature analysis and person re-identification.The specific work content is as follows:In pedestrian detection,two improved methods based on Faster R-CNN network are proposed.The first method integrates SENet(Squeeze-and-Excitation Networks)to weight features.Firstly,feature maps are generated by extracting pedestrian features through backbone network(VGG16 network)and fed into the RPN network(RegionProposal Network).Candidate regions are obtained by moving anchor boxes.Then,the generated feature maps and candidate regions are converted into fixed-size outputs by pooling operation.Finally,the loss function determines whether the candidate box is a pedestrian based on the output of the pooling layer.Experimental results on the Caltech data set show that the miss rate of the improved method with SENet is 0.999% lower than that of the original method without SENet.The second method integrates the GN(Group Norm,GN)module to normalize the features.The experimental process of this method is the same as that of the first improved method.Experimental results on the Caltech data set show that the miss rate of the improved method with GN module is 0.665% lower than that of the original method without GN module.In feature analysis,the visualization of pedestrian features based on HOG feature and Faster R-CNN is implemented respectively.The pedestrian feature extracted by the two methods was analyzed and compared,and the feature extracted by Faster R-CNN was more abundant.In person re-identification,by introducing center loss,a person re-identification method based on softmax loss and center loss joint supervision training is proposed.First,the Market-1501 data set is used to train the ResNet-50 network before and after improvement;then,the pedestrian features in the dataset were extracted using the trained model;Finally,the metric method of Euclidean distance,the metric method for the fusion of Euclidean distance and Jaccard distance are used to calculate the characteristic distances of different pedestrians respectively,and then the final person re-identification results are obtained by sorting the distances.The experimental results on the Market-1501 data set show that for the metric method using Euclidean distance,the method of adding center loss improved by 0.18%,0.89%,and 1.86% on rank1,rank5,and mAP(mean Average Precision),respectively;For the metric method of fusion distance of Euclidean distance and Jaccard distance,the method of adding center loss has improved 0.63%,0.95%,and1.25% on rank1,rank5,and mAP,respectively.In the realization of person re-identification software,Matlab2016 b is used as the development platform and Caffe is used as the deep learning framework.The person re-identification software is programmed and tested by using data sets and images collected in real scenes.The software has functions such as pedestrian detection,feature visualization and pedestrian re-identification.