Research On Pedestrian Re-Identification Method Based On Multi-Scale And Attention Enhancement

In recent years,with the improvement of safety awareness,people’s request for public safety continues to rise,and a large number of intelligent monitoring devices are deployed in crowded places such as streets and stations.However,when it is necessary to screen and discriminate monitored video sequences,it is very expensive to rely solely on human resources to seek effective data from the vast amount of data with extremely low value density.With the development of computer vision,pedestrian recognition has gradually become an important component of intelligent monitoring,which improves recognition accuracy while reducing labor costs.However,in real monitoring scenarios,pedestrian re recognition technology still faces many challenges.Aiming at the problem that existing pedestrian recognition methods cannot extract robust pedestrian features due to uncontrollable factors such as occlusion and angle changes,which affects the recognition accuracy of the model,this paper designs a multiscale feature pedestrian recognition network based on frequency-domain strip pooling and unified aggregation gates.In addition,in order to better suppress the interference of background information and extract more robust pedestrian features,this paper proposes a dual channel attention pedestrian recognition feature extraction method.The specific research content is as follows:(1)Construct a multi-scale feature model based on frequency-domain strip pooling and unified aggregation gatesAiming at the problems of feature interference between similar pedestrians,insufficient global feature representation,and lack of fine-grained information in pedestrian recognition,this paper constructs a multi-scale feature model based on frequency-domain strip pooling and unified aggregation gates.Firstly,a frequency-domain strip pooling module is designed in the model,which increases the receptive field of feature extraction through discrete cosine transform,fully extracting horizontal and vertical context information,avoiding the problem of important feature loss caused by background information interference.Secondly,a unified aggregation gate mechanism is introduced to dynamically combine output features of different scales,assigning different weights to feature maps of different scales,and further optimizing feature maps of different scales.This method extracts pedestrian features from different depths of the backbone network,better preserving pedestrian feature information at different scales,and enhancing the fine grained representation of global pedestrian features.Finally,through simulation experiments,compared to other mainstream multiscale models,FSP-UG has significantly improved on Rank-1.(2)A dual channel attention model based on MLP weighting and spatial transformation was constructedAiming at the problem of variable pedestrian attributes,this paper constructs a dual channel attention model based on MLP weighting and spatial transformation.DA-MLPSNet adaptively calibrates channel responses from both channel and spatial dimensions by combining MLP weighted channel attention and spatial transformation attention,allocating spatial weights,and making the model focus on invariant key features of pedestrian attributes.The MLP weighted attention module achieves channel level feature fusion by weighting channel feature information.In addition,combined with the spatial transformation attention module,feature maps are grouped in spatial dimensions and rotated in different directions,thereby fusing more feature information from different spaces.Utilize the complementarity of attention mechanisms in channel and spatial domains to enhance the ability of the model to extract more discriminative feature information.Compared with other mainstream pedestrian recognition models,DA-MLPSNet has better performance through simulation experiments.