Research On Person Re-Identification Method Based On Feature Fusing And Joint Learning

Person re-identification is an important field of computer vision and an important link in intelligent monitoring systems.Person re-identification achieves the automatic matching of the same pedestrians under different cameras with non-overlapping views by establishing the corresponding relationship between the pedestrian images of different cameras,which has very high application value.With the rapid development of deep learning,great progress has been made in person re-identification,but it still faces severe challenges.In order to improve the accuracy of person re-identification in practical scenarios,this paper conducts research on the deep learning method of this problem,and focuses on the two aspects—the network structure designing for feature extracting and the supervising of loss function for network training.In order to make the pedestrian features extracted by the convolutional neural network as comprehensive as possible,a structure called SPFF(Spatial Pyramid Feature Fusing)is proposed,which combines coarse-grained global features with fine-grained local features,and combines high-level semantic information with low-level detail information of the network feature map.In addition,with the help of carefully designed local feature branches and the complementary effects of global branches and local branches,the effect of pedestrian image misalignment can be reduced.Then,the SPFF structure is introduced into the baseline network constructed with Res Net50 as the backbone to form the SPFF-Net model.The model is trained end-to-end with cross-entropy classification loss function.The experiments in this paper are performed on the large-scale person re-identification dataset Market1501.The experimental results show that the accuracy of the SPFF-Net model with multi-branch features is greatly improved compared with the baseline model.In addition,from the perspective of supervising for network training,the SPFF-Net model is improved based on the loss function.Combining the characteristics of the model network structure,a multi-branch and multi-loss joint learning strategy is proposed.By combining the classification loss and the metric learning loss in different branches,the model considers the classification of different pedestrians and the difference of the same pedestrian from different angles corresponding to different branches.Therefore,the ability to distinguish the features extracted by the model can be improved.The random erasing data augmentation method is introduced during training to improve the generalization ability of the model,and the branch-grained triplet sampling strategy is used to sample difficult samples to ensure the convergence ability of each metric learning loss branch.The experimental results show that the accuracy of the SPFF-Net model is further improved from the perspective of network training,and the comparison with some excellent models that have been published so far shows the advancedness of the final model in this paper.