Research On Long-term Object Tracking Based On Kernelized Correlation Filters

Object tracking is a key issue in the field of machine vision,with the goal of locating the object of interest in the sequence of images stably and accurately.Object tracking has a wide range of applications in the fields of missile guidance,human-computer interaction,wearable equipment,auto-driving,video surveillance and so on.However,object tracking faces many challenges in the practical scenes.For example,changes of scale,posture and illumination,complex background,occlusion and leaving the field of view all impact the long-term tracking results.In recent years,the discriminative tracking algorithms have been widely concerned,and the object position is determined by the classifier trained online.The tracking algorithm based on kernelized correlation filters has achieved very good tracking performance,and performs a very high tracking speed to meet the real-time processing requirements.Based on the kernelized correlation filters tracking algorithm,our paper studies and alleviates the problem that the kernelized correlation filters tracking can't adapt to the object scale change and can't deal with long-term object tracking accurately from multi-view.To cope with the issue that tracking precision is decreased when the kernelized correlation filter faces the object scale changes and occlusion,this paper introduces the scale invariant feature,and keypoints tracking module to the kernelized correlation filter based approach.Moreover,we use the local saliency mechanism and stability analysis to filter the reliable keypoints in the search area,and we get the final results by weighting the two outputs.Through two independent modules complement and mutual restraint each other,the experiments show the object position accuracy is improved.For the second issue,i.e.,kernelized correlation filters tracking drifts or even fails in complex scenes,we propose a separate detector module with global saliency mechanism based on the kernelized correlation filters to re-detect object in the extreme tracking scenes,thus ensuring the ability of long-term and stable object tracking.