The Method Of Pedestrian Re-identification Based On Residual Network

Pedestrian re-identification aims to accomplish specific pedestrian matching in images or video sequences under non-overlapping view fields.However,in real scenarios,various factors such as camera view,resolution,occlusion,illumination and complex backgrounds can reduce the robustness of pedestrian features extracted by pedestrian re-identification algorithms,resulting in recognition accuracy that cannot meet the application requirements.To solve the above problems,this dissertation focuses on the pedestrian re-identification method based on residual network,and the specific work is as follows.A pedestrian re-identification method based on the channel attention mechanism is proposed to enhance the feature discrimination and model generalization ability.Based on the SCPNet network model,firstly,the channel attention mechanism named SE module is embedded in the backbone network Res Net50 to weight the key feature information for reinforcement;secondly,the dynamic activation function is used to dynamically adjust the parameters of Re LU according to the input features to enhance the nonlinear expression capability of the network model;finally,the gradient centralization algorithm is introduced into the Adam optimizer to enhance the training speed and generalization capability of the network model.The performance is evaluated on Market1501,Duke MTMC-ReID and CUHK03 datasets,and Rank-1 is improved by 2.17%,2.38% and 3.50%,and mAP is improved by 3.07%,3.39% and 4.14%,respectively.A pedestrian re-identification method based on feature association and multi-loss fusion is proposed to reduce noise interference in the occluded area.Based on the PGFA network model,firstly,the human key points generated by the pose estimator are used as auxiliary information to guide the model to focus on the unobscured regions of pedestrian images and extract pose guided global features;secondly,the global contrastive pooling module is introduced to fuse the features of average pooling and maximum pooling to extract global features that are more resistant to background noise and occlusion;then,the One-vs-rest relation module is introduced to tap the intrinsic relation of local chunked features and extract local features that can reflect the overall image information;finally,three loss functions,namely,weighted fusion cross-entropy loss,hardsample sampling triad loss and central loss,are used to supervise the model to learn pedestrian features with large inter-class distance and small intra-class distance.The performance is evaluated on the Occluded-Duke MTMC dataset,and the Rank-1 and mAP reach 54.9% and41.5%,respectively.A pedestrian re-identification method based on color space transformation and lightweight networks is proposed to narrow the gap between heterogeneous modalities.Based on the AGW network model,firstly,the visible image is converted to HSV color space,and the V component,which only describes the light and dark information of the image,is extracted and replicated and expanded into a three-channel image to reduce the reliance on color information;secondly,the three-channel image is downscaled and upscaled by a lightweight modal generator to generate intermediate modalities between visible and infrared images to narrow the inter-modal differences;finally,the three modalities are used as input to learn cross-modal information by a weight-sharing feature learner.The performance is evaluated on the SYSU-MM01 and Reg DB datasets,and Rank-1 is improved by 6.67% and 1.18%,mAP is improved by 6.47% and 1.15%,and m INP is improved by 5.59% and 0.42%,respectively.In terms of software implementation,Pytorch is used as the deep learning framework and PyQT5 is utilized for interface design,and the pedestrian re-identification software based on residual network is programmed and implemented,and validation tests are conducted using data collected in real scenarios.