Research Of Person Re-identification Method Based On Deep Features

Person Re-identification is one of the research hotspots in the field of computer vision.It plays an important role in helping public security organization track criminals' trajectory,and helping large supermarkets analyze customers' shopping intentions by guiding intelligent video surveillance to automatically match pictures of the same person from different cameras.However,there are still many challenges in Person Re-identification.It is difficult to query all the correct results at one time due to a lot of factors such as camera view and human pose.When a model with high performance is migrated to a new scene or a new domain,the accuracy will be severely reduced.This paper uses deep learning methods to extract the deep features of person images,and provides effective solutions to above problems.The main works are as follows:Firstly,in terms of supervised person re-identification,this paper proposes an improved query expansion method which integrates person image generation.Person images are generated by pose transfer with the help of channel attention mechanism.Then we have improved the existing query expansion method and added generated images to enhance the diversity of former query images.Finally,the accuracy of model is better for the correct matching results can get higher rank positions.We make experiments on Market-1501 dataset about person image generation and person re-identification,respectively.The retrieval results are 91.7% of m AP and 97.0% of R-1.Secondly,for cross-domain person re-identification,this paper proposes a joint learning method based on feature memory mechanism.We input the source domain and the target domain into the network at the same time for feature learning.Besides,according to the independence hypothesis of features,a memory bank module which is updated by running average is designed to store target domain features.The network combines loss functions of two domains to optimize the training process,in order to continuously improve the generalization and migration capabilities of the model,and achieve high performance in cross-domain person re-reidentification.A large number of experiments on multiple public datasets and a self-built dataset demonstrates the effectiveness of our proposed method.It reaches 62.2% of m AP and 76.3% of R-1 when transferring from Market-1501 dataset to Duke MTMC-re ID dataset,and 67.9% of m AP and 82.4% of R-1 when transferring from Duke MTMC-re ID dataset to Market-1501 dataset.