Research On RGB-D Occlusion And Deformation Objects Tracking Technology Based On Depth Model

As one of the most basic and active research topics in computer vision,object tracking has been widely used in artificial intelligence,video surveillance,and unmanned driving.However,the problems of appearance changes,occlusions,deformations,scale changes,illumination changes and other issues in the tracking process make reliable tracking still a difficult point in this direction.In order to achieve more accurate real-time RGB-D object tracking and robust occlusions and deformations processing,this paper proposes an object tracking method based on depth model which improves on the basis of kernelized correlation filter.The main work of this paper is as f ollows:Aiming at the expression ability of apparent model and the fusion of tracking results in the kernelized correlation filter tracking process,a multi-feature representation model and a method for the fusion of tracking results through feature confidence are proposed.In order to enhance the expressiveness and robustness of the target apparent model,the HOG features sensitive to geometric information are used in the Depth information feature extraction,and the 11 color name features with richer expression capabilities for color information are introduced in the RGB information feature extraction.Since the kernelized correlation filter is tracked separately in the RGB and Depth sequences,there is a change in the reliability of the two results when the appearance of the target changes.In order to obtain a more accurate total object position estimate,the online regression model and the adaptive relative value method are used to measure the confidence of the two results,and the results are fused with confidence.Aiming at the robustness of the model in the process of depth information segmentation and the change of the aspect ratio of the tracking bounding box when the target is occluded or deformed,the depth model based on the whole patch and the block patches and the region growing method are proposed to deal with the aspect ratio.If there are parameters set by people in the depth model establishment process,the tracker robustness will be reduced when dealing with various tracking scenarios.In order to eliminate human intervention,an overall depth model based on Gaussian mixture clustering and a depth model based on block depth histogram are proposed.In the Gaussian mixture clustering method,only the initial component number needs to be estimated.In the block depth histogram,the difference between the depth mean of each sub-block and the target depth mean of the previous frame is used to assign weights,and then the weights are allocated to achieve the acquisition of the overall target model.In view of the problem that the correlation filtering must maintain the size of the template,this paper proposes to keep the template unchanged in the frequency domain and combine the depth model parameter processing scales in the time domain.At the same time,the occlusion and deformation are processed by the region growth method.The experimental system designed in this paper works well,and the experimental results have certain precision.It proves that the proposed technical scheme and the adopted principle are correct,and it has important reference value for the development of the actual assisted descending system and the homing guidance system.