Research On Person Re-identification Technology Based On Deep Learning

Person re-identification technology is the identification and retrieval of specific pedestrians in different scenarios across cameras,and is widely used in public safety,intelligent security,and human-computer interaction.Due to the presence of posture,illumination,occlusion,resolution,and viewing angle changes in the actual scene,the appearance characteristics of pedestrians will be quite different,and the performance of person re-identification will decrease.Based on the method of deep learning,this paper proposes the following two improvement methods from the two aspects of network structure design and loss function:(1)Aiming at the problem that the extracted pedestrian image features are relatively single,by combining global features and multi-granularity local features,a pedestrian re-recognition method based on an improved multi-branch network structure is proposed.This method uses ResNet50-IBN-a as the backbone network.The multi-branch network structure is divided into Top DropBlock branch,global feature branch and two local feature branches.It can effectively extract local detailed features and global features of different granularities to obtain a more comprehensive feature representation.At the same time,the softmax loss and triplet loss function are used to train the model.(2)Aiming at the problems of pedestrian occlusion or posture change in pedestrian re-recognition,a method of pedestrian re-recognition based on the context relevance of local features is proposed by analyzing the correlation of the local features of the pedestrian image segmentation.Design the local feature context correlation strategy,combine the adjacent horizontal strips after segmentation to obtain the correlation between adjacent local features,thereby obtaining a richer feature representation.At the same time,the softmax loss,center loss and triple loss function are used to jointly train the model to further improve the classification effect and generalization ability of the model.Comparing with several state-of-the-art person re-identification methods on Market1501,DukeMTMC-reID and CHUK03 datasets,the experimental results show that our improved algorithm can effectively increase the performance of person re-identification.The improved multi-branch network structure can effectively extract the detailed features of pedestrian images and the local feature context correlation strategy can reduce the impact of occlusion or posture change.