Research On Person Re-identification Algorithm By Part Matching And Attention Models

In computer vision area,person re-identification is a challenging task as it involves large variations in human appearances,human poses,background illuminations,camera views,etc.Therefore,how to extract the distinguishable features for pedestrian matching is indispensable and critical for the person re-identification task.And with the rapid development of deep learning,the advantages of convolutional neural network provide a good technical basis for solving the person re-identification.The pedestrian feature automatically extracted by the network model can effectively improve the efficiency and accuracy of person re-identification.In this paper,based on the deep learning method,two models viewed from two different perspectives for person re-identification are studied.The details are as follows:To solve the problem that the traditional person re-identification method directly extracts the global features and it is difficult to obtain a better pedestrian discrimination,this paper proposes a novel deep Siamese person re-ID network equipped with an attention mechanism,constrained by a multi-loss function.The attention mechanism enhances the discriminability of the network by emphasizing effective features and suppressing the less useful ones.The purpose of the multi-loss function is to diminish distances of identical persons and at the same time expand distances between dissimilar persons in the learned feature space.Extensive comparative evaluations demonstrate that the proposed method significantly outperforms a number of state-of-the-art approaches,including both conventional and deep network based ones,on the challenging Market1501 and CUHK03 data sets.To solve the problem that the global features could not reflect the details of pedestrians,it is difficult to effectively improve the accuracy of pedestrian recognition,a multi-feature fusion method is proposed.This method does not rely on additional skeleton key points or attitude estimation models.It generates global features and part features through a feature division model and maintains the content consistency of the parts across different images.At the same time,an adaptive scoring network is constructed to unite person parts,and the pedestrian descriptor is extracted by a combination of the global feature and the united parts.Extensive experiments are conducted on three widely used benchmarks including Market1501,CUHK03 and DukeMTMC-reID,and results demonstrate that the proposed model achieves a significant performance against a number of state-of-the-arts.