Person Re-identification Based On Deep Feature Supplement

With the increasing investment in the field of public security,it is estimated that the number of surveillance cameras in the world will exceed 1 billion in 2021,and each city will have hundreds of thousands of cameras.Faced with the huge amount of video and image data generated every day,it is extremely inefficient to search a specific target pedestrian artificially.Therefore,there is an urgent need for an intelligent computer vision system to help searching the target pedestrian automatically and quickly.For this purpose,person re-identification,which is dedicated to matching the identity of pedestrians under different cameras,have been attracting extensive attention.Deep learning is very powerful in computer vision,and person re-identification also takes advantage of deep learning in representation ability to achieve rapid development.However,there are still many problems unsolved due to interferences such as illumination,posture and view.Moreover,in the actual scene,the pedestrians captured in the monitoring image are sometimes partially occluded by objects,which brings more severe challenges to person re-identification.Based on deep learning,this paper proposes two network models to address the problem of image quality difference and occlusion in person re-identification,which can provide effective supplementary information for deep person feature to enhance its robustness and improve the performance of person re-identification.The main work of this paper is as follows:(1)For the inconsistency of person image quality,a low resolution assisted network(LRAN)is proposed to extract and fuse deep person features.Firstly,a three branch network is established,which takes the original RGB image,gray image and low resolution image as input respectively.Then,the features of the three kinds of image are concatenated as the final person representation.In the test stage,the concatenated feature is used to calculate the similarity among person images.Because the low resolution image and gray-scale image provide important supplementary information for the original feature,the experiment of this method on several general person datasets achieves well performance.(2)For the inconsistency of feature distribution between holistic person images and partial person images,a global relational knowledge distillation(GRKD)method is proposed to force the output of partial person network to mimic the output of holistic person network.After back propagation training,the hidden global correlation knowledge can be naturally transferred from the holistic person network to the partial person network,and the feature spaces of partial persons and holistic persons tend to be consistent,so that it is easy to calculate the similarity between them.The experiment on partial person datasets show that GRKD can get higher retrieval accuracy.