Research On Correlation Filter Tracking Algorithm In Complex Scenes

As one of the fundamental research directions in the field of computer vision,target tracking technology plays an important role in many fields,such as intelligent monitoring,human-computer interaction,medical diagnosis and other fields.However,there are complex situations in practical application,such as illumination variation,scale variation,out of view and occlusion,which have brought great challenges to achieve fast and robust tracking effect.In recent years,with the continuous breakthrough of correlation filter and deep learning tracking algorithms,the performance of object tracking algorithm has been greatly improved,especially correlation filter,which solves the real-time problem of the algorithm.In view of the challenges brought about by complex scenes in tracking,two improved methods are proposed based on the kernelized correlation filter algorithm,which respectively improves the adaptability of the kernelized correlation filter tracking algorithm to complex scenes such as occlusion,out of view,and scale variation.The main research contents of this paper are as follows:Firstly,a kernelized correlation filter algorithm based on re-detection is proposed to solve the problem of easy tracking failure when targets are occluded in complex scenes.A confidence filter is trained,and its correlation response and average energy of correlation peaks are introduced to estimate the confidence,besides,the maximum response of the confidence filter is used to correct the position,which alleviates the tracking drift problem of the position filter.In view of the situation that the target is occluded or out of view in the process of target tracking in complex scenes,the support vector machine model is proposed to be used as the re-detection module,re-detecting the target with low confidence tracking results,this effectively improves the tracking performance of the algorithm in occlusion or out of view.Secondly,in view of the challenges of scale variation that are often faced in the tracking process,an adaptive scale kernelized correlation filter tracking algorithm is proposed.Taking advantage of the nature of scale scaling in the original coordinate system corresponding to the translation in the log-polar coordinate system,the algorithm constructs a scale correlation filter to implement scale adaptive tracking.Thirdly,a series of comparative experimental analyses were performed on the public datasets.And the results indicate that the re-detection-based kernelized correlation filter algorithm has a tracking success rate of 0.559 and an accurary of 0.764 on the OTB-100 dataset,which is improved by 8.6% and 7.8% compared with the kernelized correlation filter algorithm.And the tracking success rate and accuracy on the UAV20 L dataset are 10.6% and 14.8% higher than that on the kernelized correlation filtering algorithm,respectively.Besides,the success rate of adaptive scale kernelized correlation filter algorithm on the OTB-100 dataset is 0.548,and the accuracy is 0.752,which is an improvement of 7.5% and 6.6% respectively compared to the kernelized correlation filter algorithm.The improved algorithm in this paper can still track the target stably and accurately in complex scenes such as the occlusion,out of view,and the scale variation,which has very important theoretical value and application research value.