An Adaptive Target Tracking Algorithm Based On Circulant Kernel Matrices

With the constant progress of the society, the development of science and technology is playing a more and more important role in today’s society, the development of science and technology level represents the comprehensive strength of a country. Computer vision is an important research field of science and technology, It is closely related to people’s life and becomes an indispensable part of people’s life. Meanwhile, computer vision has been widely used in various fields like human-computer interaction, missile guidance, traffic safety, and it has got great concern of both academia and business. Among them, video target tracking technology is an important part of the field of computer vision, and it is the premise of why computer vision can be applied to other areas.Target tracking algorithm based on video can be classified into different categories:classic MeanShift target tracking algorithm, improved adaptive CamShift target tracking algorithm and currently proposed self-adaptive fast tracking via spatio-temporal context learning and exploit the circulant structure of tracking-by-detection with kernels. Although above algorithms have achieved good experimental results to some certain extent, they can’t be adaptive to the change of the target size.MeanShift algorithm is a kind of mean shift algorithm. It has the characteristic of simple calculation and fast calculation speed, and it is not sensitive to small amounts of shade, rotation, deformation. But the tracking window size of MeanShift algorithm cannot change as the target size changes, so when the target moves fast, the tracking target can be easily lost. CamShift algorithm has improved this deficiency of MeanShift algorithm. But, when the background color and the target color is too similar, CamShift algorithm is also prone to error tracking. The method of fast tracking via spatio-temporal context learning makes full use of the relationship between the target and the surrounding space, at the same time it update the spatial relations in time. It can track effectively even when severe occlusion has happened, and it also has fast process speed. The method of exploit the circulant structure of tracking-by-detection with kernels takes sample fully, and it has the advantage of tracking adequately, fast tracking speed and it can adapt to the condition of partially occlusions. But it can’t adapt to the changing size of target.Finally, this paper proposes an adaptive target tracking algorithm based on circulant kernel matrices. The algorithm uses the classification of the regularized least squares method which has a similar classification effect to support vector machine (SVM) model. By looking for implicit solution of least square method, it finds the final location of the target. To be specific, firstly, the algorithm manually labels the target area, and set the corresponding area as the region of interest. Secondly, it samples the region of interest as much as possible, and uses all samples to form a circulant matrix structure. Then, it forms the circulant matrix structure into a circulant kennel matrix structure by using Gaussian kernel function. Here in order to solve the problem of large sample size and high calculation complexity, the algorithm establishes the connection between mature circulant matrix and Fourier transform. In Fourier space, it realizes rapid learning and detection and achieves real-time performance. Finally, on the basis of tracking target successfully, and according to the response to target of KRLS classifier, the algorithm realizes the goal of being adaptive to the changing size of target. The experiments show that to a certain extent the algorithm can effectively achieve the goal of being adaptive to the changing size of target.