Research On Person Search Algorithms Based On Deep Learning

In recent years,with the rapid development of science and technology,the people pay more attention to safety issues while constantly improving the quality of life.Person search is a combination of person detection and person re-identification tasks,which positions and matches target pedestrians from uncropped raw images,and thus has great application prospects in the field of intelligent security.However,the current one-stage anchor-free person search model,although fast inference,is limited in its retrieval performance due to the lack of region alignment function of the two-stage person search model.In addition,person search is also faced with the influence of scale change,occlusion and other problems,which makes the extracted pedestrian features less robust,leading to a considerable number of mismatches.In this thesis,we will study the person search model with one-stage without anchor,solve the regional alignment problem of person search without anchor,and propose a solution for the occlusion problem to reduce the impact of occlusion on the model performance.Specifically:(1)For the problem of lack of region alignment of one-stage anchor-free person search model,this thesis studies the model structure and parameters.In terms of the model structure,this thesis proposes a mutual learning person search model based on region alignment,which adds the region proposal branch to the anchor-free person search model,and guides the person search model to learn the feature extraction method of region alignment through the mutual learning framework.In terms of model parameters,this thesis has designed the parameters of the region of interest network by clustering and other means.Experimental results show that the region-based mutual learning person search model can effectively learn the pedestrian feature extraction ability of region alignment,and its performance exceeds the benchmark model while keeping the model running speed constant.(2)For the occlusion problem in person search,this thesis studies the feature extraction and matching.In the feature extraction stage,this thesis puts forward the Vision Transformer(Vit)occlusion attention model,the model will each region proposal features of each spatial area information,extract the global semantic information,through the attention mechanism area of interest dynamic adjustment,focus on the pedestrian itself,reduce the interference of occlusion factors.In the matching stage,this thesis proposes a global-based pedestrian matching algorithm.Taking the pedestrians around the query pedestrian as auxiliary information and matching all pedestrians in the gallery to be inquired,improving the traditional pedestrian matching strategy.The experimental results show that the Vit occlusion attention mechanism and the global pedestrian matching algorithm effectively reduce the mismatching problem caused by occlusion.In conclusion,this thesis investigates the region alignment and occlusion problems based on the one-stage anchor-free person search model.The experimental results show that the mutual learning and occlusion attention mechanism are used to further suppress the influence of noise outside and inside the pedestrian detection box on pedestrian identity feature extraction and matching,and effectively improve the accuracy of person search.