Research On Algorithm Of Visual Tracking Algorithm Based On Constrained Correlation Filter Aided By Depth Information

As an important research branch of machine vision,visual tracking has become an upsurge because of its wide application requirement and practical significance.The correlation filter based visual tracking algorithm has attracted a lot of researchers' attention,which has advantages of speed and accuracy.Significant advances have also been achieved.However,in the actual environment,the target tracking may failed due to occlusion,target deformation and scale changes.With the development of RGB-D(RGB-Depth)sensor,depth image can reflect spatial information,providing an effective spatial basis for occlusion detection and target recovery.Therefore,the tracking algorithm based on RGB-D has gradually gained attention and development.In this paper,the advantages and disadvantages of correlation filter tracking algorithm and the application of RGB-D in target tracking are studied.Based on the previous research results,aiming at the problems of target occlusion and multi-pose scene,the research contents and methods are as follows.1)To overcome the problem of tracking failure in occlusion scene,an antiocclusion RGB-D based visual tracking algorithm with the aid of spital information is proposed.Firstly,cascade segmentation of depth targets is carried out on RGB-D by depth distribution histogram clustering and 2D connectivity judgment.At the same time,spatial information such as mask,depth distribution mean variance and occlusion coefficient are obtained.Then,the spatial information of the target is combined with the correlation filter response to obtain the accurate estimation of the occlusion status of the target and the occlusion is detected accurately.When the target in the occlusion state,the target is restored tracking from the target spital candidate points.Experiments show that the proposed anti-occlusion algorithm can accurately detect occlusion and restore target tracking.It also improve the anti-occlusion performance of visual tracking algorithm.2)To deal with the problem of the tracking failure in large-scale target deformation and multi-pose conversion scenes,a precise spatial constraint method of correlation filter based on constraint correlation filter visual tracking algorithm is proposed,which can eliminate edge effect and break through the limitation of partial tracking at same time.Further,a multi-model adaptive updating method is proposed to solve the problem of tracking failure when a single model can not accurately describe the target in the scene of large-scale deformation and multi-pose conversion.