Research On Algorithms Of Person Re-identification Based On Convolutional Neural Network

Person re-identification is a technology that retrieves and matches target pedestrians across cameras and scenes.This technology has been widely used in intelligent video surveillance,smart commerce,and intelligent security,and is also a important research topic in the field of computer vision in recent years.However,due to the effects of changes in perspective,pedestrian occlusion,and pedestrian pose changes in natural scenes,person re-identification faces major challenges.How to extract more discriminative features for person matching is still the focus of current person re-identification concerns.In this paper,based on deep learning methods,we study the improvement of the feature representation ability of person re-identification models.The main research work of the paper is as follows:(1)Existing methods of person re-identification based on convolutional neural network lack discriminative information caused by occlusion and complex backgrounds.In order to solve this problem,a method based on multi-scale convolutional feature fusion is proposed.In the training phase,to extract multiple eigenvectors containing global features and multi-scale local features,we use the pyramid pooling method to block and pool the convolutional feature map.Then we classify each feature vector independently and normalize the weights and features on the last inner layer of each class to improve the classification performance.Finally,we use the gradient descent algorithm to optimize the sum of losses for each classification.In the recognition phase,we concatenate the pooled multiple feature vectors into a new vector for similarity matching.we verify the efficiency of the proposed algorithm on datasets,Market1501 and DukeMTMC-reID,The results show that the proposed algorithm can effectively improve the accuracy of person re-identification.(2)Existing methods of person re-identification based on local feature have weak ability to discriminate features,due to pedestrian misalignment and pedestrian pose changes in the detection bound box.To solve these problems,a multi-task pyramid overlapping matching pedestrian feature method is proposed herein.In the training phase,the feature map extracted by the backbone network is used to obtain a global vector and a plurality of local vectors with different scale features through the pyramid overlapping matching network,These vectors are jointly learned using three loss functions,and a feature normalization layer is used to reduce the impact between loss inconsistent learning targets to extract more appropriate shared feature.In the inference phase,the dimensionality-reduced feature vectors of each branch are concatenated into a new feature vector for similarity matching.This method is used for comparison experiments on the Market1501 and DukeMTMC-reID datasets,and the results show that the features extracted by the model have strong robustness and discriminative power.(3)Aiming at the problems of complex model structure based on local features and the lack of discriminative semantic information for low-level person re-identification methods,a person re-identification method based on batch feature erasure and multiple feature fusion was proposed.Firstly,low-level detail features and high-level semantic features are extracted through the residual network,and low-level features are used to compensate for the lost detailed information of high-level features.Second,batch random erasure is used for high-level semantic features to obtain discriminative local features.Finally,use multiple loss functions for supervised learning of three features.During the test,three features are fused into a feature vector for similarity matching.Experimental verification on the datasets Market1501 and CUHK03 shows that the method can not only improve the accuracy of person re-identification,but also reduce the calculation amount of the model and increase the speed of the algorithm.