Research And Implementation Of Multi-object Detection And Tracking Algorithm In Expressway Service District

Along with the growth of expressway and service district,the number of video surveillance system is increasing.Due to the increasing number of travellers,traditional video surveillance system could not meet the demand of security.It is of great research significance to replace traditional manual methods with artificial intelligence technology in analyzing and processing surveillance video data.Service district intelligent monitoring system is an important part of both intelligent surveillance video system and expressway intelligent management monitoring system.In this thesis,multi-object tracking algorithm for pedestrians and vehicles which is based on object detection is realized.From the real service area surveillance video,images with different perspectives,different heights and distances,and different backgrounds are selected to label,and the pedestrians and vehicles detection data set is made.On the data set,the sate of art target detection algorithm is optimized and trained.By comparing the actual detection speed and accuracy of these algorithms,YOLOv3 is set as the final object detection module in this thesis.Object re-identification can be used in multi-target tracking algorithm.Robust and effective feature descriptors can improve the recognition rate.In this thesis,the residual network structure of the cosine metric learning network is optimized by using the splittransform-merge structure and the parameter amount is reduced by 20% end up with a lightweight appearance feature extraction model.A vehicles and pedestrians re-identification data set is made to solve the problem of multi-class target re-identification in surveillance video.Both two cosine metric learning networks were trained and tested on the Market1501,MARS and the self-built re-identification datasets to verify the validity of the model.In this thesis,a multi-target tracking framework based on detection is implemented.Then the trackers are initialized and the Kalman filter is used to predict the state of each target.The feature extraction network trained in this thesis extract the target features and construct the correlation matrix to match the tracker and the detection target.The association forms an associated trajectory.Finally,the algorithm is compared and verified on the surveillance video target tracking data set made by the virtual reality and multimedia technology laboratory of Southwest Jiaotong University...