Research On Person Re-identification Based On Adaptive Feature Distance Fusion

As an important research branch of video processing,person re-identification aims to solve the problem of person retrieve in cross-scenes.However,due to the complexity of surveillance scenes,camera view changes,person pose changes,and occlusion,it has become a challenging subject.In order to effectively reduce the impact of various disadvantages and further improve the accuracy of re-identification,this paper relies on deep learning to design an adaptive weighted feature distance fusion network based on feature contribution,and design a level feature network with joint loss to optimize feature extraction.Besides,this paper design a multi-strategy cooperation re-ranking method to conduct research on person re-identification.The main research work includes:(1)An adaptive weighted feature distance fusion network based on feature contribution for person re-identification is designed.Divided into multi-stream feature extraction network and adaptive multi-stream feature distance fusion network.The multi-stream feature extraction network can extract image global feature,and aligns local areas of person body based on pose estimation module to extract local feature.To reduce background influence,image segmentation and person re-identification are combined to introduce mask local feature.Adaptive multi-stream feature distance fusion network combines global feature and local feature,considering the importance of different feature,propose the concept of feature contribution,and utilize weight decision module to adaptively assign weights,combined with the feature distance to obtain the final distance as image similarity descriptor.This method achieved Rank-1accuracy of 92.3% and 84.7% on CUHK03 dataset and Market1501 dataset.(2)A level feature network with joint loss is designed to optimize feature extraction process.Level feature network considers the limitations of top-level feature,combine low-feature with top-level feature to enhance feature expression capabilities.In addition,the level feature network combines metric learning with person re-identification,utilizes Tri Hard loss to constrain the low-level feature learning,and proposes a hybrid loss function Mix loss to constrain the top-level feature learning,thereby improving the network generalization ability.This method can optimize the feature extraction process,the accuracy of Rank-1 is increased to 93.4% and 85.7% on CUHK03 dataset and Market1501 dataset.(3)A multi-strategy cooperation person re-identification re-ranking method is designed.This method is based on the aggregation of the neighborhood match ranking method,the bidirectional search ranking method,and the reciprocal neighbor ranking method.The neighborhood match ranking method constructs image neighborhood according to feature distance,and calculates image similarity based on the distribution of the neighborhood set in the ranking list.The bidirectional search ranking method combines the forward and backward ranking list,considering the mutual constraint relationship between image pairs.The reciprocal neighbor ranking method improves the Jaccard distance and combines it with the original feature distance to design a new metric distance.In this paper,three sub-methods are aggregated into multi-strategic cooperation re-ranking method,which combines similarities of multiple categories,re-ranks the initial ranking list,put image with same labels on the top of the ranking table and remove the wrong matching image as much as possible.The accuracy of Rank-1 is increased to 95.7% and 88.9% on CUHK03 dataset and Market 1501 dataset.