Research On Person Re-identification Algorithm Based On Deep Learning

With the accelerating process of urbanization,a large number of surveillance video data have been produced in the field of intelligent security.Person re-identification,as an important method to analyze these video data,has attracted extensive attention of scholars at home and abroad.However,there are still many difficulties in person re-identification,such as different camera angles,large background differences,pedestrian posture changes and occlusion.Therefore,this paper studies the person re-identification task through the method of deep learning.The main work is as follows:Firstly,aiming at the problem that the detailed information of image will be lost when convolutional neural network extracts features,a global multiscale pooling network(GMPN)is proposed.The model combines the maximum pooling operation with the average pooling operation,which retains the most significant information in the local area without damaging the integrity of the original information;Different receptive fields are used to enrich the features extracted by neural network;The features of different scales are fused by Hadamard product,and the differences between the two features are preserved while retaining the important information of the two features at the same time.Secondly,aiming at the insufficient utilization of the features extracted in different stages of network in neural network,a multiscale feature stitching network(MSFS)is proposed.The model is divided into one main branch and three sub branches.In the sub branches,multi-scale pooling is used to aggregate the information in the feature map,and the attention mechanism is applied to extract the global features of different scales.Thirdly,aiming at the lack of using second-order statistical information to model attention model,a second-order information and relation-aware attention network(SORAN)is proposed.The model uses point product and second-order statistical information to model the correlation between features,and combined with the idea of relation-aware to mine the global structural information,the network can pay attention to more important places.Finally,the three methods are verified by experiments on Market1501,Duke MTMCre ID,CUHK03 and MSMT17 datasets,and m AP,Rank-1,Rank-5 and Rank-10 are used as evaluation indexes.Compared with other person re-identification algorithms,the effectiveness and feasibility of the proposed method are proved.