Research On Person Re-identification By Fusing Global And Local Features

Person re-identification refers to the tracking of the same individual across multiple surveillance cameras and is an important application in video surveillance.However,due to significant variations in the pose,viewpoint,and occlusions of persons in different cameras,accurately identifying the same individual is a challenging task.Global feature extraction aims to capture the overall information of persons,while local feature extraction improves sensitivity to local parts,thereby mitigating issues such as pose variations and occlusions.In this thesis,we study the fusion of global and local features in person re-identification algorithms,focusing on extracting global and local features.The specific contributions are as follows:(1)To address the problems of accuracy degradation caused by the loss of critical person features or pose variations,we propose a person re-identification model that combines global and local feature branches.The model consists of two branches: the global branch extracts global features using a multi-receptive field fusion module to capture contextual information,followed by GeM pooling to obtain fine-grained features.The local branch extracts local features using a PCB network that evenly divides the global features into six local features.The input image is first aligned and enhanced using a spatial transformation network and then fused with the global and local feature maps using a pixel-level attention network.Finally,an improved triplet loss and cross-entropy loss are combined in a specific ratio as a multi-loss joint function to supervise the training of the network.Experimental results on the Market-1501 and DuckMTMC-reID datasets demonstrate mAP values of 86.2% and78.7%,respectively,proving that this method can improve the accuracy of person re-identification.(2)To address the problem of suppression of local features during training of convolutional neural networks,we propose a person re-identification model that incorporates multiple branches for feature fusion and learning.In order to enhance the learning of local attention features and weak feature parts,an erasure branch is added to the global and local branches.To preserve image edge information,we optimize the method of extracting local features by employing a step-wise approach to segment the feature map,strengthening the connections between local features.Subsequently,a multi-scale adaptive attention module is used to obtain well-fused local feature maps.Finally,all the feature vectors from the global branch,local branch,and erasure branch are concatenated to form the final person image feature vector.Supervised training is performed using a label-smoothed cross-entropy loss function.Experimental results on the Market-1501 and DuckMTMC-reID datasets demonstrate mAP values of 86.5% and 78.9%,respectively,demonstrating that this model can capture more details in person images.(3)Based on the aforementioned algorithm research and the person re-identification workflow,we propose a prototype system for person re-identification and design the overall system architecture.According to the requirements analysis,the system implements person detection and person image matching functionalities.The system has a user-friendly interface and simple operation,proving its practical value.