Research On Visual Object Tracking Method Based On Discriminative Correlation Filter

Achieving stable and accurate object tracking in complex scenarios is an urgent focus issue in the field of visual object tracking.In recent years,Discriminative Correlation filters(DCF)-based visual target tracking algorithms have attracted much attention due to their advantages in balancing tracking speed and accuracy.Nevertheless,the performance of traditional DCF is not satisfactory for the challenges of occlusion,background clutter,and illumination changes that commonly exist in complex tracking scenes.This is due to the following reasons: 1)the lack of ability to suppress response distortion;2)the boundary effects caused by cyclic assumption;3)the model degradation caused by linear update strategy.To address the above problems,this thesis will investigate and improve the traditional correlation filters from the following aspects,in order to improve the anti-interference capability and finally achieve stable and accurate tracking performance.(1)To address the problem of filter response distortion due to background interference,this work proposed a background-suppressed dual correlation filter for visual tracking.The overall strategy of adopting the target response to constrain the fluctuation of the background response is implemented to achieve the purpose of limiting the variation rate of the global response and to realize the effective suppression of the background interference.First,the mask matrix is used to crop the target features from the global features.Second,the spatial regularity constraint and the background response suppression regularity are introduced to construct the dual regression model to train the target filter and the global filter respectively.The differences between the two output response maps are exploited for mutual constraint,so as to highlight the target and suppress the background interference.In the detection process,the global response map is enhanced by weighted fusion of the target response maps,which further enhances the tracking performance in complex scenes.Experimental results on the OTB100,TC128 and UAVDT benchmarks show that the proposed algorithm achieves an average accuracy of 79.3% and average success rate of 68.5%,which is an improvement of 7.1% and 6.5% respectively compared to the baseline BACF tracker.(2)In order to mitigate the boundary effects and model degradation problems,this work proposed an adaptive spatial regularization correlation filter with augmented memory for visual tracking.By imposing an adaptive spatial regularity to penalize unreliable non-target regions,the target features are effectively highlighted while mitigating the boundary effect.Secondly,to alleviate the model degradation problem,the filter is co-trained by selecting some historical views together with the current appearance model,thus achieving enhanced memory effect.In addition,to further improve the anti-interference ability of the model,this work introduces the aberrance repression regularization to limit the drastic changes in the response maps while establishing a high-confidence model update strategy for maintaining a relatively pure training sample model in the complex tracking environment.Specifically,two metrics,maximum response peak and average peak fluctuation,are adopted to measure the tracking quality.Based on this,the learning rate of the appearance model is dynamically adjusted to reduce the over-learning of unreliable target appearance,so as to construct a robust training sample model.Experimental results on the OTB100,TC128,UAVDT and UAV123 benchmarks show that the proposed algorithm achieves an average accuracy of 76.0% and average success rate of 75.5%,which is an improvement of 5.3% and 5.1% respectively compared to the baseline BACF tracker.