Research On Person Re-Identification Based On Contour Information Embedding

In recent years,the urbanization construction of our country has reached a new height,and the video surveillance system can be seen everywhere in the city has become an important tool to protect people’s life and property safety.Person re-identification is one of the important roles of video surveillance.Its fundamental task is to use the surveillance system that has been built to complete cross-camera pedestrian tracking,and accurately find the target person in the cross-domain and cross-time surveillance video.Person re-identification technology is an important research direction in the field of machine vision,which is a sub-problem of image retrieval.Nowadays,deep learning has been widely used in the field of image processing and machine vision.Similarly,in person re-identification research,the method based on convolutional neural network is also the most mainstream research direction.Existing research results show that convolutional neural networks are more inclined to obtain texture-based features than shape-based features in the process of feature extraction.In addition,some researchers have used pedestrian contours in a feature-aided manner in re-identification systems..Based on the above existing research results and findings,this paper proposes the following innovations to solve the problem of lack of contour information in person reidentification research using convolutional neural networks:Firstly,a person re-identification method based on contour image embedding is proposed.The existing edge extraction model is used to extract the contour of the commonly used public data set,and the pedestrian contour map data set is constructed.Since it is difficult to directly change the feature extraction logic of the convolutional neural network,it is proposed to use the pedestrian contour map as a feature supplement to add to the middle layer of the network,so as to increase the amount of contour information in the feature extraction process.This paper verifies the best embedding model on the backbone network Res Net-50 and improves the recognition ability of the network through a large number of experiments.Secondly,a person re-identification method based on contour information embedding is proposed.In the process of experimenting with the first method,this paper finds that adding contour features artificially in the network will cause a certain amount of information waste,and the best way to deal with it is to let the network learn more contour information by itself.Based on the above reasons,this paper proposes a contour information extraction module according to the attention idea.Its input is the contour feature map and the middle layer feature map of the backbone network,and the relationship matrix is constructed according to the correlation between the two,and the feature map with more contour information bias is sent to the rest of the convolutional neural network to continue training,so that the network can learn more contour information.In this paper,the proposed person re-identification method is tested on the commonly used public datasets Market1501 and Duke MTMC-re ID.The experimental results show that the person re-identification method based on contour image embedding outperforms the baseline model by 3.5% in m AP and 3.4% in Rank-1,and can be applied to other baseline models without increasing the complexity of the model.The person re-identification method based on contour information embedding has a m AP of 83.8% and a Rank-1 of 95.1% on the Market1501 dataset,and a m AP of 73.5%and a Rank-1 of 86.8% on the Duke MTMC-re ID dataset.Compared with other advanced methods,the recognition effect of this method is still competitive.In summary,the method proposed in this paper makes the convolutional neural network have more contour information bias,so as to improve the final recognition ability of the network.It is a breakthrough attempt in the use of contour information in the field of person reidentification,and has important significance for the subsequent person reidentification research.