Research On Pedestrian Re-identification Method With Multiple Visual Semantic Information Embedding

Pedestrian re-identification(Re ID)has been applied in varies fields such as intelligent surveillance and smart cities.In recent years,the most Re ID methods use the visual semantic information of body parts to guide the network learning local features of pedestrian.Nevertheless,the images of Re ID mainly come from the video surveillance of real scenes which leads to the occlusion and low resolution.In these cases,the human contour information is an important visual semantic information.In order to enhance the applicability of Re ID methods in real scenes,this paper proposes to use multiple body parts and human contour information to form the multiple visual semantic information,and then research and design a Re ID method with multiple visual semantic information embedding.The mainly research work in this paper includes:(1)In order to obtain visual semantic information which includes multiple body parts and human contour information,a body parts semantic segmenter is designed and implemented to obtain visual semantic information of multiple body parts.The comparison experiment results of this segmenter on the LIP data set show that the segmenter can accurately divide the image foreground into multiple body parts and obtain their information.Then,some main deep learning-based contour detection methods are analyzed and compared on the BSDS500 data set,which show that the HED model has a better performance.Therefore,the HED model is selected as the human contour detector to obtain the visual semantic information of the human contour.(2)In order to extract the discriminative features of pedestrians in multiple scenarios,a Re ID method with multiple visual semantic information embedding is proposed.This method has a multi-branch structure,the main branch extracts global features and the other branches respectively embed the multiple body parts and the human contour information obtained by the segmenter and detector into the high-level features to extract the corresponding local features.For the problem that high dimension of concatenate features,this paper uses a fully connected layer to fuse the local features of various body parts.Experiments on the proposed method are conducted on benchmark data sets and the results have verified that the multiple visual semantic information composed by multiple body parts and human contour information can improve the accuracy of Re ID.(3)Further research on the proposed Re ID method with multiple visual semantic information embedding has found that the problems of insufficient representation of pedestrian features and too many parameters in the fully connected layer possibly exist.Therefore,this paper proposes a further improvement method to deal with above problems.The improved method embeds multiple visual semantic information on low-level features to obtain local features,and then each local features is abstracted through the network,finally uses a convolution layer instead of the fully connected layer to fuse multiple local features.The improved method has been tested on multiple data sets,and the results show that embedding multiple visual semantic information on low-level features and using convolution layer fusion features can improve the accuracy of Re ID.(4)Based on the previous research,a Re ID prototype system with multiple visual semantic information embedding is developed which taking the finally given Re ID method with multiple visual semantic information embedding as the core.This paper focuses on the Re ID method with multiple visual semantic information embedding.This method guides the network learning local features of pedestrian through multiple body parts and human contour information to enhances the application of Re ID method in low resolution,complex backgrounds and different pedestrian scales.Based on the previous research,a Re ID prototype system finally is developed.