Research On Problems Of Person Re-identification Based On Deep Learning

In recent years,due to the large application significance in intelligent surveillance system,the person re-identification(re-id)problem has become one of research focus in the computer vision field.The aim of person re-id is to determine whether the pedestrian captured by one camera is the same person as a pedestrian taken by another non-overlapping camera at different times.At present,the person re-id problem can be generally divided into three categories:image-based person re-id,video-based person re-id and scene-image based person search.In image-based person re-id,one or a few images of each person under each camera are employed to solve the problem.While the video-based person re-id adopt a series of video sequences of each person,which contains extra temporal motion information.In the scene image based person search task,in addition to solving the person re-id problem,the person detection,i.e.,the localization of target person in the entire scene image,is also involved.There are several difficulties in the person re-id:the distraction of illumination variance and occlusions,and the inaccuracy of person detection.In this work,considering the characteristics of above three tasks separately,the deep learning based person re-id system is build.The main research contents and contributions of this work are as follows:1.Image-based person re-id.Based on the biological characteristics of human visual system,this thesis proposes an end-to-end comparative Attention Network(CAN)model.This model exploits a three-branch deep neural network framework to simulate the human visual attention mechanism to compare and search for multiple local regions of person image pairs.Finally,the regional features are extracted and integrated to improve the recognition rate of image-based person re-id.2.Video-based pedestrian re-id.In order to fully utilize the temporal motion information contained in the person video sequence,an end-to-end two-stream deep neural network,called Accumulative MOtion Context Networks(AMOC),is proposed in this thesis.The model can automatically learn the appearance-motion features from raw video sequences,and continuously accumulates the effective spatiotemporal information in a recurrent way.Finally,the disctiminative person features are generated.The analytic experiments on three benchmark datasets validate the effectiveness of AMOC.3.Scene image based person search.Scene image based person search is a new problem in the person re-id field.This thesis utilizes the shrinking attention mechanism and coins a deep model that can joitly solve the person re-id and detection problems,which is called Neural Person Search Machines(NPSM).The NPSM has a unique memory unit regarding the target pedestrian as a memory.Under the guidance of memory,the searching region of interest can gradually recursively shrink from the entire scene image to the target person area.The experimental results show that NPSM can well perform the shrinking process of visual attention,and improve the accuracy of person search.