Research On Anti-drift Visual Tracking With Double Enhancement Of Information

Visual tracking usually take the feature of tracking objects as the tracking information for model training and detection.In addtion,the tracker only can get accurate information from target detection or artificial tagging in the first frame,and the information used to update the tracking model is provided by the predicted results in subsequent frames.Both the quantity and quality of the tracking information are seriously deficient.Once incorrect prediction or severe occlusion occurs,the tracking model will drift.In allution to the above problems,based on the theory of discriminant tracking,an anti-drift tracking algorithm with double enhancement of information is proposed.The single feature representation of target appearance is extended to the multi-information fusion representation of the appearance,context and motion state of target.The sub-models of correlation filter,color probability and optical flow motion are conducted,where the above information is extracted and calculated separately and the response map about the center position of target will be outputted.Then the weighted sum of all response maps is defined as the total response map to achieve the fast integration of all sub-models and information.Meanwhile,according to the peak value and the fluctuation of the total response map,the confidence of predicted result in each frame is evaluated.Only when the result satisfies the high confidence condition,the new target information will be learned and the model will be updated.The experiment results show that the improved method of double enhancement of tracking information is feasible and effective.The expansion of information quantity can enhance the representation ability of the tracker and the evaluation of information quality can reduce the pollution on tracker caused by incorrect prediction.The combination of these two approaches can effectively solve the problem of model drift in the process of tracking and improve the accuracy and robustness of the tracker.Compared with the benchmark algorithm only based on the information of target appearance,the proposed algorithm improves the average distance precision by 5.6% and the average success rate by 4.7% on the OTB-2015 data set.And the average success rate of the proposed tracker is improved by 10.2%,9.6% and 7.4% respectively in the video sequences with the problems of out-of-view,low resolution and motion blur.The proposed algorithm can achieve the same robust tracking as the trackers adopting deep feature or complex classifier under the condition of simple feature representation and classifier.There are 24 figures,9 tables and 61 references in this paper.