Research On Person Re-identification Method Irrelevant To Clothing Change

Most of the conventional person re-identification(Re-ID)datasets are collected in a short time span,during which the clothes and appearance of people hardly change.However,in many real-life scenarios such as shopping mall monitoring or criminal tracking,the same person’s clothes may be changed,and different people may wear similar clothes.Conventional person re-identification methods rely too much on person’s clothes information to match them,which is unsuitable for clothing change scenes.This paper explores the problem of person Re-ID under clothing change,and proposes a feature decoupling method based on deep convolutional generative adversarial networks to separate clothes appearance features and human body structural features to extract person features that are irrelevant to clothing change.We define two encoding modules to encode the appearance and structural features of a person.The generator combines the two features to generate a reconstructed image.The feature extraction of the two encoding modules is constrained by the loss function to achieve the decoupling of appearance and structural features,and the structural features are regarded as features irrelevant to clothing change.Based on data augmentation for image hue adjustment and human parsing,two improved methods are proposed on the basis of the original method,aiming to optimize the feature extraction effect of the structural feature encoder.In addition,based on the Market1501 dataset,we use GAN to generate a Re-ID dataset Market1501＿CC for testing phase,in which clothes of people are changing randomly.Employing rank-k accuracy and m AP as evaluation metrics,experiments are conducted on the PRCC and Market1501＿CC datasets and we compare the results with other person Re-ID approaches.The experiment results show that the two improved person structure feature extraction methods in this paper perform better in person Re-ID task under clothing change.Furthermore,our two improved methods remain effective in person Re-ID task in same clothing.We apply our model in a person tracking system,and conduct experiments on test videos collected in actual scenes,demonstrating the advantage of our approach for both person tracking task under clothing change and conventional person tracking task and verifying the significance and feasibility of our approach in practical engineering applications.