Person Re-identification System In Port Based On Jetson

With the development of biometric technology,frame-based person re-identification technology has achieved excellent results.In the frame-based person detection and person re-identification tasks,a large number of low-value normal samples will be generated,making it difficult for the system to obtain meaningful samples.In this thesis,based on edge computing device,an improved person re-identification network based on mixed domain attention mechanism and multiple loss functions is proposed to improve the comprehensive performance of person re-identification network.Then,based on the Kuhn-Munkres algorithm and the idea of fusion feature center,the feature matching and feature extraction of multi pedestrians are carried out to realize the unconscious person re-identification system.The main work of this thesis is as follows:1.Propose an improved re-identification network based on mixed domain attention mechanism and multiple loss functions.For the limited memory and computing power on the edge computing device and the real-time requirements of the project,Using a kind of global person re-identification baseline and embedding a hybrid domain attention mechanism block,at the same time a set of multiple loss functions are proposed to assist the network for learning,which improves the performance of all aspects of the network with a small computational cost.Finally,the experiment on Market1501 and Duke MTMC data sets show that the improved network has advantages in speed and accuracy compared with baseline and other commonly used person re-identification network,especially on the more difficult Duke MTMC data set.2.Propose an unconscious person feature extraction based on Kuhn-Munkres algorithm and feature center.The algorithm obtains a large number of uncorrelated persons features through person detection network and person re-identification network,and then the Kuhn-Munkres algorithm is used to correlate the features between adjacent frames and gradient descent algorithm is used to calculate the feature center generation of each identity and update the features of original gallery set.Implements the unconscious person feature extraction and fusion.The experimental results show that the feature matching based on Kuhn-Munkres algorithm can effectively improve the matching accuracy,and the feature fusion method based on gradient descent can further improve the accuracy of retrieval.3.Design and implement the unconscious person re-identification system based on Jetson TX2.Based on the above research content and the proposed algorithm as the core function module,Mobile Net V2-YOLOV3 is used as the target detection network.Moreover,the feature search engine is used as the database,and a person re-identification system is designed and implemented.It realizes the re-identification of the specific stream of people.Finally,the application value of the system is verified through the system demonstration.