Research On Person Re-identification Method Based On Attention Mechanism

Person re-identification is a technology that aims to identify a target person from images or videos captured by multiple non-overlapping cameras.It has significant applications in video surveillance and criminal investigation.However,pose variations,occlusion,and cluttered backgrounds can cause significant intra-class variation,making it challenging to extract discriminative features.Attention mechanism is a technique that can enhance the discriminative power of pedestrian features.Its core idea is to weight and strengthen the key information in the feature map to reduce the interference caused by irrelevant information.Additionally,Attention mechanism can adaptively learn the focus points for different scenes,thereby improving the model’s generalization ability and robustness.This paper focuses on the application of attention mechanism in person re-identification,and the specific work is as follows:1、To address the problem of the model’s difficulty in extracting key features due to occlusion and pose variations in person re-identification,we propose a person re-identification algorithm driven by a global attention mechanism and a relation network.First,in the backbone network,a global attention module is embedded into the Res Net50 network to capture the weight information of the spatial and channel dimensions.Second,in the relation network,multiple local features at different scales are obtained through horizontal segmentation.Furthermore,two modules are designed: one for extracting global contrast features and the other for extracting relation features between local features.Then,the local relation features are fused with global contrast information and sent to the classification network.In the loss optimization stage,a joint loss is used for network training.Finally,experiments are conducted on five commonly used datasets,and the experimental results are analyzed.The Rank-1 results on the Occluded-Duke and Market1501 datasets reached 63.2%and 95.4%,respectively,and the m AP reached 53.8% and 88.2%,respectively,demonstrating the advanced nature of the algorithm.2、To verify the effectiveness of self-attention mechanism in improving the performance of person re-identification task,non-local attention and axial attention modules are applied to the person re-identification model,and experiments are conducted on the Market1501 dataset.The results show that when the self-attention module is inserted into the convolutional layer of the re-identification network,the accuracy of the model is improved.Additionally,ablation experiments show that when these two self-attention modules are used together,the performance of the model is even better.Furthermore,considering the relatively small size of the dataset,we propose a person re-identification method called NLAA-Net with a gating mechanism,which simultaneously captures global and local contextual information,which is crucial for accurate person re-identification.Experimental results validate the superiority of the proposed method and the effectiveness of the gating mechanism.3、We propose a new method called PT-Net to address the problem of performance degradation caused by occlusion in person re-identification.This method combines pose estimation and Transformer technology.First,an existing pose estimation method is used to detect keypoints in the input image,and this information is combined with person feature maps to generate a pose-based person feature representation.Then,the Transformer model is used to encode the pose-based person feature representation to achieve feature alignment and fusion.The experimental results on the Occluded-Duke dataset demonstrate that PT-Net outperforms the baseline model,with improvements of 1.3 and 1.5 percentage points in m AP and Rank-1 metrics,respectively,highlighting the superiority of this approach.