Research On Pedestrian Re-identification Method Based On Deep Learning

Pedestrian re-identification aims to judge whether pedestrians have appeared in different camera fields of view,which is a sub-problem of image retrieval.With the wide application of deep learning in the field of computer vision,pedestrian re-recognition based on deep learning has become the mainstream model,and its effect is far beyond the pedestrian re-recognition scheme based on traditional machine learning.However,in actual application scenarios,the task of pedestrian re-identification is still very challenging due to a series of problems such as low camera resolution,posture changes,frequent illumination changes,and serious occlusion.How to extract a distinguishable feature representation that is robust to illumination changes,posture changes,and occlusion is particularly important.In view of the fact that global features cannot eliminate interference information such as background,and cannot obtain more fine-grained local representations,current mainstream models use attention mechanisms to model global and local features as the final pedestrian representation to suppress interference information such as background and obtain More finegrained local representation.This paper proposes corresponding improvement schemes for the shortcomings caused by the attention mechanism and the segmentation part respectively,and the main overviews are as follows:Aiming at the problem of insufficient ability of the attention mechanism to extract potential salient features,this paper proposes a Cascaded Attention Enhancement /Suppression Network(CAE/SN)based on the attention mechanism.The salient attention is enhanced in the current stage and suppressed in the next stage to extract potential salient features.Finally,multi-level salient regions are excavated and a robust feature representation is obtained.In view of the fact that the model based on horizontal segmentation to obtain multi granularity local features can not solve the problem of pedestrian part misalignment,this paper abandons the simple horizontal segmentation method,introduces the clustering algorithm to generate pseudo tags as the semantic supervision information of pedestrian image,and uses pixel level pseudo tags to determine the local area and obtain multi granularity local features.In this paper,we combine the two improvements and propose a cascade attention guided by clustering(CAGBC)network,which uses semantic pseudo tags to guide the division of salient regions in the stage of significant attention enhancement / suppression,and gets rid of the problem of insufficient ability of attention mechanism to extract potential features and semantic misalignment caused by horizontal segmentation,The multi-level robust salient features are obtained.The model is tested on two large-scale datasets of market-1501 and dukemtmmc.The results show that the model achieves 88.7% and 80.1% in map evaluation index respectively,and it is 1.1% higher than DSA-reid model with attitude assistance on Market-1501 datasets.This shows that our model achieves the same results as the model using additional semantic pre training without using additional semantic information(such as gesture detection and human semantic parsing),and further illustrates the significance of extracting multi-level salient features and guiding salient region selection with semantic pseudo tags.