Kernel Learning Estimation: A Model-Free Approach to Tracking Randomly Moving Object

Abstract

Kernel learning estimation (KLE) is a kernel-based method, where the original spatial data is mapped into a high-dimensional Hilbert space by a nonlinear map**, hiding the nonlinear map** in a linear learning framework. The kernel function of the method can be used to replace the complex inner product operation in the high-dimensional space and avoid the Curse of Dimensionality caused by high-dimensional calculation effectively. The kernel-based method has advantages on learnability, computational complexity, precise linearization and generalization performances, providing a promising way to solve the problem of nonlinear target tracking. In traditional tracking methods, nonlinear tracking models are usually built as a priori to predict the current state of target motion, emphasizing on tracking accuracy and real-time performance. However, kernel-based method provides a general way of linearization processing, which can be independent of specific models to achieve highly efficient data-driven computation. Introducing the kernel learning mechanism into target tracking problem is expected to improve the environmental adaptability. In this paper, a review on kernel learning method with application to randomly moving target tracking is presented, including kernel-based algorithms for target detection, kernel-based algorithms for generative tracking and for discriminant tracking, and multi-kernel learning methods with multiple kernel functions. Further research is prospected in optimization of kernel function, long-term robust tracking, feature extraction, target occlusion and other potential aspects on moving target tracking using kernel learning theory.