Research On Person Re-identification Based On Single-scale Features And Attention Mechanism

Person Re-Identification is a computer vision task that aims to retrieve the provided query person from non-overlapping multiple camera views of pedestrian images using computer vision techniques.Person re-identification has a wide range of applications and can be combined with pedestrian detection technology to be used in areas such as video surveillance,intelligent security,and so on.However,in current practical applications,person re-identification technology still faces many problems,such as difficulty in obtaining annotated pedestrian data and the problem of pedestrian occlusion by objects or unrelated pedestrians,which make the task of person re-identification quite challenging.To address the above problems and improve the performance of person re-identification algorithms in practical applications,this paper designs an unsupervised person re-identification algorithm based on soft labels and an occlusion person re-identification algorithm based on single scale,respectively.The main work of the paper includes the following three aspects:1.To address the difficulty in obtaining annotated data,this paper proposes an unsupervised person re-identification framework based on soft labels.Traditional unsupervised methods mostly use clustering to generate pseudo labels,but simple clustering often leads to hard quantization loss.To avoid the impact of hard quantization loss,this paper proposes using a Memory Bank to store non-parametric person features.In each iteration,the similarity between the sample features and the person features in the Memory Bank is obtained,sorted according to the similarity,and the most similar features are dynamically assigned soft labels using Euclidean distance to train the network.As this method assigns soft labels based on the similarity between features,a fusion attention mechanism is also proposed to improve the robustness of the features.By combining channel attention and self-attention mechanisms,the feature representation ability of the person re-identification model is jointly enhanced.2.To address the problem of occlusion in person re-identification,this paper proposes a person re-identification model based on single-scale features.To enhance the robustness of single-scale features,a data preprocessing method with occlusion data augmentation is proposed.Additionally,a spatial memory module is introduced to model the input feature map and the key-value pair memory network to remove occlusion interference and obtain more discriminative person features.Finally,we found that information from the same camera view is not conducive to model learning.Therefore,a penalty term is added to negative samples from the same camera view based on the triplet loss to alleviate the interference of information from the same camera view.Experimental results demonstrate that the model based on single-scale features can also obtain more discriminative person features.3.This article utilizes PyQt to design a visual program interface.By combining pedestrian detection model and person re-identification model,the program achieves the effect of pedestrian positioning and tracking.In the visual program interface,all pedestrians in the video and the target pedestrian to be searched can be detected in real time.