A G2G Similarity Guided Pedestrian Re-identification Algorithm

As an important intelligent video analysis technology,Person re-identification(Re-ID)is widely used in intelligent security,case detection,lost search,intelligent interaction and other fields.It is an important means in the field of public security and has become a research hotspot of many scientific research institutions.However,the lighting,occlusion,resolution,background and posture of human body in the actual scene are different,and the shortage of datasets makes the person re-identification task still faces many difficulties and challenges.This paper focuses on the robust pedestrian feature extraction in complex environments and the methods to improve the utilization of data resources.More efficient and accurate recognition results have been obtained.The specific contents are as follows:Based on the Siamese Network structure,this paper improves and optimizes the backbone ResNet-50 for pedestrian feature extraction,so as to solve the problem that the video surveillance scene is complex,the person image resolution is low,and it is difficult to extract stable pedestrian image features.On the one hand,the ordinary convolution in the model is improved based on the idea of depth-separable convolution,the convolution process of channel feature and spatial feature is separated,which realized the uncoupling of channel and space,and the reduce of operational pressure.On the other hand,according to the characteristics of channel features to images,a bilinear channel fusion attention mechanism is designed to fuse the image features under different receptive fields and convert them into the attention weights of different channel features,so as to fuse and enhance the fine-grained feature information of different channels and improve the robustness of features.This paper proposes the idea of optimizing the similarity between the Probe and the Gallery(P2G similarity)by the similarity between images in Gallery(G2G similarity)and has given a specific algorithm to solve the problem of the less person reidentification data resources and low utilization rate.Based on the idea of random walk,the G2G similarity is used to influence the initial P2G similarity score through the random walk model,so as to realize the fine-tuning and correction of the initial P2G similarity,which not only leads to the optimization of P2G similarity,but also brings abundant monitoring signals for network training.At the same time,the difficult sample mining module is designed,and the more difficult sample is selected for similarity optimization,so as to enhance the generalization ability of the model and reduce the calculation pressure of the network.In addition,the horizontal overlapping grouping of image features was used to calculate and optimize the global and local P2G similarity separately,and the weight of each group's features was learned through the fusion training process to realize the automatic perception of the importance of each key part.In this paper,the Siamese Network structure and the similarity measure updating method are improved.And the experimental results show that the proposed method,compared with the existing method,can effectively improve the accuracy of person re-identification task.