Research On Person Re-Identification Method Based On Multi-Granularity Feature Fusion And Local Attention

As the country accelerates the process of urban intelligence and informatization,the field of intelligent video surveillance continues to expand.In recent years,computer vision technology has been mainly used to process and analyze monitoring videos,and person re-identification is one of the important research directions.It is commonly used to solve cross-camera tracking and monitoring security issues.Person re-identification is a retrieval task that matches a person across cameras based on their appearance,posture,and other features given a person image.Due to the complexity of the scene in the pedestrian image captured by monitoring cameras,there are various factors such as occlusion,illumination changes,background changes,and posture changes that interfere with the image,resulting in a large amount of noise and affecting the useful information of the person,ultimately leading to a low recognition accuracy.Therefore,how to effectively extract the features of person,enhance the model’s robustness to adverse factors such as occlusion,and improve the accuracy of person re-identification is a challenging and meaningful task.With the development of deep learning,most existing methods for person re-identification focus on learning discriminative and robust features through feature embedding.Despite achieving some progress,there is still substantial room for improvement in terms of person recognition accuracy.This thesis focuses on the following work for person re-identification:This thesis proposes a network model based on random occlusion and multigranularity feature fusion to address the problem of person re-identification in complex scenes,which involves occlusion and monotonicity in person discriminative features.Firstly,the input image is processed by random occlusion to simulate the real-life occlusion scenarios and enhance the network’s robustness to occlusion.Secondly,a multi-granularity feature extraction network is designed,consisting of a global branch,a local coarse-grained mutual fusion branch,and a local fine-grained mutual fusion branch,to extract global holistic features and supplement local deep features with multiple granularities to enrich the hierarchical discriminative features.Then,the local information fusion module is employed to further strengthen the correlation between local features.Finally,the network is trained with a combination of triplet loss and label smoothing loss functions.Performance evaluation of the proposed method is conducted on three standard public datasets,namely Market1501,DukeMTMC-re ID,CUHK03,and an occlusion dataset called Occluded-Duke.The experimental results show that the proposed method achieves a Rank-1 accuracy of up to 95.2%,demonstrating its effectiveness in improving the precision of person recognition.This thesis proposes a network model based on local attention and saliency mining to address the problems of insufficient attention to local features and neglect of potentially salient features in pedestrian re-identification.The aim is to improve the network’s ability to acquire and process person features,thereby enhancing the discriminative power of the model.Firstly,based on a multi-granularity feature extraction network,a local attention branch is proposed to increase attention to local features using channel attention mechanism and hybrid pooling module,capturing highly distinctive local features and reducing the impact of occlusion and background interference.Secondly,through a local saliency mining branch,potential salient features behind the focused features are extracted,which is beneficial for enhancing the discriminative and representative power of local features and improving the robustness of the network in complex scenes.Comparative experiments are conducted on multiple standard public datasets and an occlusion dataset.The results demonstrate the superiority of the proposed method,with the highest mean Average Precision(mAP)reaching 87.9%.This showcases improved recognition performance and confirms that the features obtained by the proposed method are more discriminative.