Research On Person Re-Identification Based On Feature Disentangling And Feature Fusion

Pedestrian re-identification is a crucial area of research in computer vision,as it finds wide applications in intelligent video surveillance and intelligent security with the increasing number of surveillance cameras.Pedestrian re-identification aims to establish matching relationships within multiple cameras with non-overlapping shooting areas,and it is a sub-problem of pedestrian retrieval.Pedestrians have features that are multi-layered and strongly coupled,and different shooting locations or times can lead to large differences in lighting,angles,postures,backgrounds,etc.in pedestrian images.Pedestrian re-identification techniques based on deep learning,which extract deep networks with rich and complex features,have developed rapidly in recent years.However,researchers still face challenges in this area.This study proposes a GAN-based unsupervised pedestrian re-identification framework that analyzes and improves the performance.First,the framework decouples the identity and structural features of the pedestrian.The structural features are extracted using a 3D human pose and shape estimation network,and the original features are rotated to obtain a new structure with multiple angles.Second,the identity features are fused with the new structural features to obtain the generated images of the pedestrian from different angles.Additionally,the model introduces a comparison module to enhance the online data and improve the generation quality by learning features with view invariance.(1)First,this research proposes a novel approach to address the multi-layered and strongly coupled features of pedestrians,which includes decoupling two different levels of features: identity features and structural features.A 3D human pose and shape estimation network is used to extract the structural features of pedestrians,which are then combined with the identity features to generate images of the pedestrians from multiple angles.This increases the size of the dataset and improves the generalization capability of the model.Additionally,a comparison module is introduced to enhance the online data and improve the quality of the generated images.Ablation experimental results demonstrate the effectiveness of the comparison module,indicating the potential of the proposed approach in pedestrian re-identification tasks.(2)Second,this study proposes a method to improve the metric learning component of pedestrian re-recognition.This method is based on adaptive sparse genetic algorithm,which is used to optimize the initial ranking list and improve the accuracy of rerecognition.(3)Thirdly,this study introduces a fuzzy filtering module to select high-quality generated images.Typically,generated images lack viewpoint-specific information,which results in low-quality images.The fuzzy filtering module can extract the fuzziness of the generated images and filter out the low-quality images that are not suitable for network training.By integrating the fuzzy filtering module and the reordering algorithm into the system,the model’s burden is reduced,and the noise brought by low-quality generated images is eliminated,leading to improved system accuracy.