Research On Single Object Tracking Of Visual Target Under Complex Scenario

Visual tracking is an important research area in computer vision.It has been widely applied in video surveillance,traffic control,motion analysis,et al.The goal of visual tracking is to track the specific moving object in videos.Interference factors like object occlusion,illumination variation,pose variation,scale variation and motion blur always appear in the real complex scenario,it makes accurately and effectively visual tracking become difficult.This thesis aimed at analysing and researching the single visual object tracking problem to design robust visual tracking algorithms with efforts.Aiming at the partial occlusion and tracking drift problems which exist in the target tracking process under complex scenario,a new visual tracking algorithm was proposed based on the combination of sparse representation and spatial-temporal context information.In this algorithm,a new visual tracing algorithm framework based on multiple templates matching was constructed,and the appearance information of targets in the initial,previous and historical frames was fully utilized.A patch based sparse representation was adopted to handle the partial occlusion,and the spatial-temporal context information was organically combined in the multiple templates matching framework with a new mechanism.This algorithm can effectively handle the partial occlusion problem,meanwhile,it obviously improved the tracking drift problem.The tracking accuarcy under the scenario with partial occlusion was improved as the spatial-temporal context information and the multiple templates weighting matching were adopted in the algorithm framework.Aiming at the problem that the features extracted by traditional methods have poor robustness and generalization ability,a new generative tracking algorithm based on deep features was proposed.According to the features of visual tracking,a deep convolutional neural network more suitable for target tracking was constructed.Then,with this network,a new multiple templates matching tracing framework based on deep features was constructed.In addition,a new adaptive update mechanism of the network and templates was constructed,the multiple templates also combined in an adaptive weighting manner.The robustness of the algorithm was obviously enhanced,the tracking accuarcy under several kinds of complex scenario was improved obviously.The algorithm also obtained good tracking performance for the long-term tracking.Aiming at the mismatching problem existed in the generative tracking methods and the problem caused by the binary-value labels existed in the discriminative tracking methods,a new generative and discriminative mechanism combined tracking algorithm based on deep features was proposed.A new tracking framework adaptive classification decision was constructed.The framework not only contained generative and discriminative modules,it also combined long-term tracking and short-term tracking in this two kinds of modules.According to the potential relations between the classification results of candidates and the appearance variation degree of the targets' appearance,a new multiple modules adaptively weighted combining mechanism was constructed.Moreover,to alleviate the problem caused by the binary-value labels,a new regression prediction network was constructed.The generalization ability of the algorithm improved obviously,and the tracing accuracy of the algorithm under different kinds of complex scenes was also improved.