Research On Cross Domain Person Re-identification Based On Spatial Local Features Interaction

As a key link in an intelligent video surveillance network,person re-identification technology aims to find the matching pedestrian image across multiple non-overlapping cameras by giving a pedestrian image,which can quickly and effectively retrieve and track a specific pedestrian in a large-scale database.This paper studies how to extract discriminative features and cross domain person re-identification by using deep learning technology.First of all,aiming at the decline of accuracy caused by cross domain person re-identification,a person re-identification algorithm based on migrating mutual accumulation learning network and Jaccard distance order optimization is proposed.The clustering algorithm is used to generate hard pseudo labels in the target domain.In order to alleviate the influence of noise labels in hard pseudo labels on the model,the classification result of hard pseudo labels as soft pseudo labels is proposed.Then,the model becomes robust by alternately performing online optimization of soft pseudo labels and offline optimization of hard pseudo labels;the re-ranking method plays an important role in improving the retrieval accuracy of person re-identification.Therefore,the Jaccard distance order optimization— k-reciprocal nearest neighbor method is used to re-rank and optimize the person re-identification results,which can effectively improve its accuracy.Secondly,in order to solve the problem that the spatial context between parts is ignored when learning local features,features fusion based on temporal learning for person re-identification is proposed.We first use the Long short term memory network based on Recurrent Neural Network to model the pedestrian in an end-to-end way,which regards the pedestrian as a sequence of body parts from head to foot,and then the more discriminative features are extracted by using the complementary information between local and global features.However,when using methods based on convolutional neural networks to extract features,some detail clues will gradually disappear with the increase of convolution layers.Therefore,the module of Channel Interaction and Integration is introduced to aggregate channel features,so as to improve the ability of features.Finally,aiming at the local parts dislocation of two pedestrian images,a person re-identification based on attention mechanism is proposed.When calculating the local distances,the dynamic programming method is used to automatically align the local parts,so as to find the shortest path of the two images.This algorithm has achieved good performance in labeled datasets,but its performance will decline in new fields.Therefore,considering that different camera configurations will change intra-domain image style,camstyle model is proposed to transfer the style of the image in the target domain to achieve camera invariance.Based on the prior knowledge that the source domain images and the target domain images can form natural negative pairs,we can directly use these negative pairs to learn the triplet loss and realize domain connectedness.