Person Retrieval Method Based On Part Features

In recent years,due to the development of deep learning,more progress has been made in pedestrian retrieval.It has good applications in many fields,such as security and safety,intelligent human search system in large public places,unmanned supermarket,home intelligent robot and so on.However,different scenes mean that the pedestrian images detected by the camera have complex differences.There are problems such as occlusion and different backgrounds in the pedestrian images.How to reduce the problems caused by various factors and improve the recognition accuracy of pedestrian retrieval is a problem that needs to be solved.To avoid the influence of the complexity of pedestrian image and improve the accuracy of pedestrian retrieval,this paper mainly carries out the following works :(1)in order to solve the segmentation method of local parts,a pedestrian retrieval method for fine segmentation of parts is proposed by extracting the features of parts.Through fine segmentation parts,the consistency inside parts is improved,the impact of hard segmentation of pedestrian images is reduced,and the retrieval accuracy is improved.(2)the multi-branch network is adopted to integrate the features of components and global features to improve the robustness of the model.(3)a self-supervised method of significant feature weight learning is proposed to enable the model to distinguish whether the feature is significant or not.When extracting features,we focus on the significant features,reduce the weight of the insignificant features,and reject the insignificant part features.In this paper,experiments were conducted on two large data sets,Market-1501 and Dukem MTMC-re ID,and the experimental results showed that the pedestrian retrieval accuracy was better than the original part segmentation method.After using the refined segmentation method,the Rank-1 reached 93.8% and 82.5%,respectively,and the m AP reached 79.2% and 62.9%.After the addition of significant feature weight learning method,Rank-1 reached 92.3% and 80.3%,and m AP reached 79.2% and62.9%.Experimental results show that the two methods proposed in this paper both improve the accuracy of pedestrian retrieval.