| Single target tracking is one of the most urgent problems in the field of computer vision,whose aim is to track the subsequent motion of the target at any given time.Tracking technology is often used in civil fields such as virtual reality and Skynet.Although the single target tracking algorithm has been developed rapidly over a recent period of time.However,in practical applications,because the target is in a complex and changeable environment,such as affected by illumination,deformation,scale,occlusion and so on,tracking algorithm is difficult to locate the target accurately.The main reason is that the error training and updating of the model weakens the ability of the tracker to recognize the target,which leads to the drift phenomenon and finally leads to the tracking failure.Aiming at the above problems,this paper mainly studies the single target tracking algorithm from the aspects of feature selection and fusion,discrimination under occlusion state,and the selection and application of classifier.The main tasks are as follows:The classical visual single target tracking method usually uses single feature to describe the target being tracked.However,in the actual scene,the apparent features of the target are affected by illumination,deformation,scale and other factors.In order to describe the target better,HOG feature and CN feature are introduced in this paper.The traditional feature extraction method is used to train their correlation filter models,and their response graphs are obtained by correlation filtering with their respective features;Then the weights of their response values are obtained by using the fixed weight method,the Peak to Sidelobe Ratio（PSR）method,the difference between the actual response and the expected response,and the sensitive value of PSR;Finally,the final response value and target position are obtained by fusion of each response graph with weights,and the model is updated adaptively.When the target is occluded,because most of the apparent features of the target are lost,it is difficult to track the target stably by the strategy of feature fusion alone.Therefore,the loss function?1?2-is introduced in the tracking module to reduce the sensitivity of the correlation filtering algorithm to local occlusion;then the PSR value of the current frame response diagram is used to determine whether the target is seriously occluded.When the target has serious occlusion,the candidate target enters the re-detection module,and then the candidate target position is detected according to the SVM classifier.Finally,the target response value of the detection module and the target response value of the tracking module are compared to the final location of the target,the classifier and the model are updated respectively.In this thesis,34 color video frame sequences in OTB2013 dataset are selected to compare and compare the algorithms,and the experimental results are analyzed qualitatively and quantitatively.The experimental results show that the tracking performance of this algorithm is better than that of other contrast algorithms While ensuring real-time performance.Therefore,the multi-feature fusion and occlusion judgment strategy play an important role in improving the performance of the target tracking algorithm. |
