| Since 2016,"Smart City" has developed rapidly in China.In order to comply with the development of smart cities,city surveillance will become more intelligent and high-definition.There are several problems exists in non-smart cameras,for example,they cannot provide real-time warning and consume a large amount of resources for analysis,it is necessary to solve those problems with the help of computer vision technologies.As a key issue in the field of computer vision,the single-object tracking algorithm can plot the trajectory of a given object to facilitate subsequent behavior analysis and anomaly detection.Therefore,as a basic problem in the field of computer vision,the single-object tracking problem has important theoretical significance and research value.In the urban surveillance scenario,the monitoring images is very complex,which mainly reflected in:(1)the type of objects is various and the density of people is very high,meanwhile,there are frequent occlusion among objects;(2)deformation and scale changes occur when objects move in surveillance video.These problems bring a lot of challenges to the single-object tracking algorithm:(1)in a complex background,the accuracy of the algorithm would be interfered with the objects which is similar to the target;(2)the temporal information of video is not fully utilized when tracking objects.This thesis studies on how to solve those challenges.The main work and innovations are as follows:1.A single object tracking algorithm based on feature fusion is proposed to solve the problem of how to reduce the interference caused by similar background under com-plex background.This thesis integrates the features extracted from different layers so that they can maintain discriminative ability and appropriate size.In order to solve the problem that the target cannot be tracked after being occluded,this thesis combine the local detection model with the global detection model:the local detection model cal-culates the accurate size and position of the object;when the object is occluded or the tracking fails,a global detection model is used to detect the possible position of the tar-get through the entire image,if the target is found,the local detection model is used to continue to track base on the position.The experiments show that the proposed algo-rithm can greatly improve the performance of the single-object tracking on the chosen dataset.2.A single-object tracking algorithm based on inter-frame information is proposed to solve the problem of how to make full use of temporal information under complex background.This thesis proposed a single-object tracking algorithm based on inter-frame information:in the sampling stage,this thesis first calculate the optical flow information between the previous frame and the current frame,and use the optical flow and the position of the object calculated in the previous frame to predict the possible position of the current frame,and use this as a reference point,this can improve the quality of the sampling.In addition,this thesis uses the target that has been tracked successfully as a sample to provide reference for the current frame.This thesis uses the Siamese network to calculate the similarity between the samples of the current frame and the sample that has been tracked successfully,and selects the most similar sample as the final tracking results.The experiments show that the proposed algorithm become more robust due to the fully utilized temporal information.This thesis is funded by the National Natural Science Foundation of China,"Co-operative Video Surveillance with Hybrid Camera Networks for Traffic Management"(ID:61371192),and studies on the single-object tracking algorithm based on deep learn-ing in a complex scenario.This thesis solves the problem of how to reduce the inter-ference caused by similar backgrounds and how to make full use of temporal informa-tion under complex backgrounds,and improves the performance and robustness of the single-object tracking algorithm. |
