Deep Learning-based Methods For Person Re-identification

With the development of social informatization,more and more large public places such as shopping malls,subways,campuses and housing estates have installed large-scale cameras to form a large-scale real-time video surveillance network.Traditional video analysis uses manual query and processing of video information to consume a lot of human and material resources.How to effectively and quickly process surveillance video has become a major problem.person re-identification,as an important link in the video surveillance network,has received more and more attention.Pedestrians in realistic and complex environments are subject to light intensity,variable pedestrian postures,motion occlusion and other issues,which lead to large differences in the collected pedestrian images,which affects the accuracy of person re-identification.By applying deep learning to the person re-identification algorithm,the recognition accuracy of person re-identification is improved,and the surveillance video information can be processed quickly and effectively,which has certain practical significance.The existing person re-identification data set has few samples and deep network layers,which is prone to network overfitting problems when training the network.At the same time,due to real-world lighting changes,camera angle changes,occlusion and other problems,person re-identification accuracy is not high,a person re-identification feature extraction method based on attention mechanism is proposed.First,the random erase method is used to perform data enhancement on the input pedestrian image;then the saliency of spatial pixel features is enhanced by constructing a top-down attention mechanism network;finally,the salient features of the entire pedestrian are similarly measured and sorted to obtain pedestrians The accuracy of re-identification.Experiments show that person re-identification based on the attention mechanism reaches 88.53% in Market1501 data set Rank1 and 70.70%m AP;in Duke MTMC-re ID data set Rank1 reaches 77.33% and m AP 59.47%.Therefore,the proposed method has a significant performance improvement on the two major personre-identification data sets,and the algorithm has good robustness.Aiming at the difficulty of extracting the features of pedestrian recognition by the existing person re-identification methods,a person re-identification method combining multi-attention network and multi-stage features is proposed.First,the spatial attention network and the channel attention network are fused to construct a multi-attention residual network;then the three-stage pedestrian features of the multi-attention residual network are extracted,and their characteristics are fused to obtain the prominent features of the pedestrian,to compensate for the lack of pedestrian information;Finally,pedestrian characteristics will be obtained for similarity measurement and sorting to obtain the accuracy of person re-identification.Experiments show that this multi-attention network and multi-stage feature fusion person re-identification method has a Rank1 of 92.62% and a m AP of 80.59% on the Market1501 dataset;a Rank1 of 86.20% and a m AP of 68.49% on the Duke MTMC-re ID dataset.The proposed method has significantly improved performance on the two major person re-identification data sets,and has certain application value.