Design And Implementation Of Pedestrian Detection And Re-identification System Based On Deep Learning

The massive data generated by the Internet not only involves simple operations such as storage and query,while how to extract the needs from big data Information is also essential.Especially in the field of security,it is a critical clue for the crimina l police to search for target person from massive video data.However,manual search is time-consuming and laborious,and sometimes key information is lost,which is inefficient.Therefore,pedestrian detection and pedestrian re-identification technology has always been a research hotspot.The pedestrian detection and pedestrian re-identification system designed in this paper is a practical project of Zhengzhou Xinda Advanced Technology Research Institute.Based on the actual demand analysis of the system,the overall architecture design and the network architecture design,by using the pedestrian detection and feature extraction algorithm based on deep learning,the Euclidean distance between the feature vector of the target pedestrian and the feature vectors of all pedestrians in the database is calculated,so as to judge the similarity of pedestrians and realize the function of fast pedestrian detection and pedestrian recognition.The main contents of this article are as follows:Firstly,we retrained and tested the detection algorithm of YOLO v3 for a single pedestrian category.The data set is composed of 5000 pedestrian images in the actual scene marked by Label Img and pedestrian images in PASCAL-VOC2012.After training,the m AP value of the YOLO v3 network model reached 92.5%,the time of single prediction was stable at 21 ms,the accuracy and speed could meet the actual needs of the system.Secondly,the Res Net50 network was retrained and tested using the Market-1501 data set.The m AP of the Res Net50 network which had been retrained was 75.57%,the probability of Rank-1 was 90.02%,the time of single image classification was stable at 28 ms.This showed that the trained network had better classification accuracy and speed for pedestrian pictures.Thirdly,the system database was designed in detail.We selected postgre SQL database which could calculate Euclidean distance quickly,designed E-R diagram of database,model diagram of all entities,logical structure of table and detailed content of table,completed and realized the function of system database.Fourthly,We tested and evaluated the practical system.On the basis of testing the basic functions of the system,the accuracy and real-time performance of the system were evaluated.In the return results of the first four pages of the system,the probability of the same pedestrian captured by different angle cameras was 90%.This also showed that the pedestrian feature vector extracted by the trained resnet50 network had a good expression ability.The algorithm took a total of 49 ms for a single image,which could achieve real-time detection and recognition while meeting the accuracy.The designed system is simple and convenient.The users can not only find the suspect through the system,but also see the pedestrian pictures that have recently appeared in the monitoring on the single or multi-channel display interface of the system,which ensures the safety in the monitoring area to a certain extent and fully meets the actual needs of Party A.The designed system has been actually applied in Zhengzhou Xinda Advanced Technology Research Institute.