| As a crucial research avenue within the field of computer vision,person reidentification offers indispensable information support for multiple tasks,such as multicamera multi-target tracking and pedestrian search.In real-world scenarios,pedestrian images are easily subject to various factors,leading to continual alterations in the image cue scale for determining whether the images correspond to the same individual.Consequently,person re-identification models still confront the challenge of inadequate capacity to capture multi-scale image cues.Furthermore,pedestrian images acquired via cameras frequently entail diverse occlusions,which can give rise to issues such as partial invisibility of pedestrian images and feature misalignment,thereby constraining the performance of pedestrian re-identification models within occlusion environments.This thesis conducts research on the aforementioned issues in the following sections:(1)To address the issue of abrupt changes in the scale of pedestrian image cues caused by factors such as viewing angle,pedestrian posture and environment,this thesis proposes a pedestrian re-identification method based on a multi-scale feature fusion Transformer.Firstly,a multi-scale feature cascade module is constructed to endow shallow image features with deep semantic information,and to obtain pedestrian feature representations at different scales,thereby enhancing the model’s capacity to capture multi-scale image cues.Secondly,scale encoding is incorporated into the feature extraction backbone to enable the model to differentiate image features from different scales.Furthermore,to improve the model’s utilization of multi-scale pedestrian local features,a feature aggregation module is designed to integrate feature information of different scales and extract effective pedestrian information from local features of varying scales.Through a series of experiments on Market-1501,DukeMTMC-reID,and MSMT17,the proposed method has demonstrated promising performance in the field of person re-identification.(2)To address the issue of occluded pedestrian images and introduction of cluttered background noise,this thesis proposes a novel method based on pose estimation and background suppression.The proposed approach leverages human pose estimation to extract local features from multiple visible parts of pedestrians as prior knowledge,aiming to reduce the impact of occlusion on the feature representation.A local feature enhancement graph convolution module is also designed to incorporate rich contextual information from global features into local features,while utilizing the pose information to model the interconnection between the local features to enhance the robustness of the pedestrian feature representation.Moreover,a background suppression module is introduced to the attention mechanism,which incorporates pedestrian pose information to guide the model to focus on the pedestrian part,thus suppressing the cluttered background noise in the pedestrian image.Experimental results on the Occluded-DukeMTMC dataset demonstrate the effectiveness of the proposed method for occluded pedestrian re-identification.(3)Incorporating the method proposed in this thesis,we integrated two pedestrian re-identification techniques with target detection and cross-camera multi-target tracking,developing a cross-camera multi-pedestrian tracking system.This system enables pedestrian search and cross-camera tracking across multiple scenes,and performs well in both occluded and unoccluded environments.The pedestrian tracking functions demonstrate high efficacy and robustness,indicating the potential of this system for real-world applications in surveillance and security. |
