Research And Implementation Of Multi-Camera Moving Object Detection And Recognition System

With the rapid development of urban and rural roads,the traffic environment has become more and more complex.Vehicle detection,tracking and re-identification technology is the key technology to realize intelligent transportation system,and has become a research hotspot in the field of computer vision.In recent years,the relevant computer processing technology has been rapidly improved,while still facing some challenges.This thesis mainly studies the tasks of multi-object tracking and re-identification,and the main work is as follows:(1)Multi-object tracking algorithm based on graph neural network.In a complex traffic environment,due to the serious occlusion of the target,there are frequent ID switching and track mismatch problems in the tracking process.To solve this problem,this thesis proposes a multiobject tracking algorithm based on graph neural network,which transforms the tracking problem into a bipartite graph matching problem between the tracked track and the current detection frame.First of all,a multi-clue feature expression learning network integrating the appearance and motion information of the target is established;Secondly,a bipartite graph structure is constructed between the tracked object and the detected object,and its fusion features are used as nodes;Finally,the graph neural network is used to iteratively aggregate the node features from the neighborhood to achieve the information transfer between nodes,and the edge regression algorithm is used to calculate the similarity matrix to achieve the tracking of the target.The network is trained by using the triple loss function and the binary cross entropy loss function.Experiments are carried out on the KITTI-tracking and MOT 17 data sets to verify the effectiveness and superiority of the algorithm.(2)Vehicle re-identification algorithm based on graph neural network.The algorithm can effectively extract all features with robustness,and take into account the spatial geometric relationship between different features,so as to better realize vehicle re-identification.For this reason,ResNet50 is adopted as the backbone network in this thesis,combining attention mechanism and multi-scale feature fusion to optimize the global feature extraction network and extract more representative global features;Then the vehicle view is analyzed to obtain four local features;Then,the graph neural network is used to study the structural relationship between these features,and the softmax loss function,center distance loss function and triple loss function are used to train the network to improve the accuracy and reliability of the model.Through comparative experiments,it is found that the algorithm shows significant effectiveness and superiority on both VehicleID and VeRi-766 data sets.(3)Research and implementation of cross-camera vehicle tracking system.Based on the multi-object tracking algorithm and vehicle reidentification algorithm studied in this thesis,cross-camera vehicle tracking is realized,and the task of matching the track ID of the same vehicle in different videos is completed.In order to reflect the application value of the algorithms studied in this thesis,PyQt5 is used to package the relevant algorithms and develop the application system,so that users can have a better use experience.