Person Re-identification Based On Convolutional Neural Network

The video surveillance has played a crucial role in city security.Given a specific pedestrian target,the task of video surveillance is localization,tracking and retrieval for the pedestrian in video sequence.The retrieval and recognition of a specific pedestrian is called person re-identification,which has been an important topic in the field of computer vision.Since images captured in surveillance are often involved scale variation,rotation,occlusions and changing illumination,person re-identification remains very challenging.Current research mainly focuses on three aspects: robust feature representations,metric learning and re-Ranking.To address the above challenge,the thesis exploits deep learning theory to investigate person re-identification methods which are applied to a single-frame image and video sequences respectively.The main contributions are given below.(1)Person re-identification based on person attributesThe current mainstream approaches for feature extraction are based on Convolutional Neural Network(CNNs).While more semantic information is contained in deep features,some detail information is lost.To tackle this issue,a person attributes based person re-identification method is proposed,where the id information and attribute information are combined to train CNNs.The overall system is composed of three parts.The first part is a basic CNN model used to extract pedestrian features.The second part,a re-ID network,employs triplet loss function and feature fusion strategy to obtain robust features.The third part is an attribute network where person attributes and cross-entropy loss are leveraged to supervise training stage.Experiments show that by training the above three sub-networks jointly,more robust features are obtained and performance of person re-identification has been improved effectively.(2)Person re-identification based on Spatial and Temporal attentionComparing with the image,video sequences contain more information of pedestrians.How to effectively utilize spatial and temporal information is the focus in the research of video based person re-identification.To this end,a person reidentification method is presented based on spatial and temporal attention.Specifically,to extract more robust features,a spatial attention model is designed to select regions from each frame and yield different attention for each region.On the other hand,a temporal attention network mechanism is adopted during feature fusion,which assigns different weights for each frame in the video.The temporal attention mechanism could alleviate the influence generated by frames with low quality.The experimental results demonstrate that the proposed spatial and temporal attention based method achieves competitive recognition performance when compared with the state-of-the-art approaches.