Research On Time-regularized Nuclear Correlation Filter Tracking Algorithm

Visual target tracking is the basic technology in the field of computer vision,which has been used in many fields in modern times,such as drones,medical image processing,video surveillance,autonomous driving,robotics,etc.Because the training data of visual target tracking is generally only the first frame,and there are often target deformation,masking,background clutter,moving out of view,etc.,the visual target tracking research is still challenging.The visual tracking algorithm based on the correlation filter is the mainstream tracking paradigm in recent years.According to the convolution theorem,the circular convolution is equal to the multiplication of elements in the frequency domain,so the algorithm has efficient operation efficiency and better tracking performance.Based on the correlation filter tracking algorithm can be divided into two categories: linear correlation filter tracker and nuclear-related filter.At present,most of the relevant filter trackers pick the update strategy is essentially self-regression model,easy to make the tracker over-fitting the current frame.If the current frame is obscured,background clutter,etc.,the filter can easily update its parameters by mistake,which may cause tracing to fail.In view of this,this thesis takes the nuclear-related filter as the research object.First designs a time regularization suitable for nuclear-related filter,so that the nuclear filter can adaptively update CFs with the latest sample and maintain the balance with the previously learned CFs.Next,the weighted APCE update strategy is designed to improve the tracking effect of the nuclear-related filter.Finally,based on the strategy of time regularization and weighted APCE update,a nuclear-related filter tracker algorithm that integrates time regularization and scale adaptation is proposed.On the classical visual tracking data set OTB,the comparison experiment with the existing nuclear-related filter tracker algorithm,and the three angles of quantitative results of overall tracking performance,quantitative tracking performance results of sequence properties,and qualitative tracking performance results,illustrate the validity and robustness of this method.