RGB-D Object Tracking Based On Multi-features Kernel Correlation Filter

Object tracking is popular for the current research area,and it is also a difficulty research that attracts much attention in the field of computer vision.With the development of modern society,it has a very wide range of applications,including intelligent city,artificial intelligence,video surveillance,and other fields.In the past recent years,a lot of excellent works have been found in the related research.However,due to the influence of local occlusion,scale transformation,illumination change and so on,it is still difficult to achieve reliable tracking of universal targets in unconstrained environment.A lot of research work is carried out from two aspects of the tracking model and the apparent model,trying to solve the difficult problems in the study of the above algorithms.In order to solve the difficulties in the object tracking of RGB-D,based on the kernel correlation filtering theory,this paper has improved the robustness of the object tracking by using the apparent model,and proposed an improved adaptive multi-features tracking model for the defects of the multi feature fusion kernel filter.The main research work of this paper is as follows:(1)The improvement of the apparent model: in order to adapt to the illumination changes,morphological changes and visual angle changes of the target in the tracking process,this paper introduces multiple features to improve the performance of tracking.In this paper,four complementary features are extracted about grayscale,color and depth map to improve the robustness of target representation.The robust Color Name based on color image extraction,Enhance Histograms of Oriented Gradients based on gray image extraction;based on depth value extraction for occlusion robust Cross-bin Distribution Field.Deep edge extraction is for Depth Histograms of Oriented Gradients.The experimental results show that the multi-features apparent model has higher perform than the single feature model.(2)The improvement of the tracking model: in order to solve the problem of the dimensional inconsistency between the kernel related filters for multi-features fusion,this paper proposes a multi-features tracking model improvement.On the basis of depth scaling kernelised correlation filters,the problem of missing feature dimension in multiple feature fusion of nuclear related filters in DSKCF is solved.The multi-features tracking model avoids the lack of multi dimension by extending the label dimension.In addition,in view of the shortcomings of the original model,the model updating strategy is divided into response update and autocorrelation update,and the independent update strategy is used to improve the robustness of model updating.Sufficient experimental data show that the multi-features tracking model brings further performance improvement on the basis of the existing multi-feature appearance models.