A Review on Kernel Learning Method of Moving Target Tracking

Abstract

The kernel method maps the original spatial data to a high-dimensional Hilbert space by nonlinear mapping and hides the mapping in the linear learner. The kernel function is used to replace the complex inner product operation in high-dimensional space, which can effectively avoid the ‘curse of dimensionality’ caused by high-dimensional space calculation. The kernel method has the advantages of learnability, efficient calculation, linearization and good generalization performance, which provides a new effective way to solve the problem of nonlinear target tracking. The traditional target tracking methods often use the tracking model to predict the current motion state of the target and ensure the accuracy and real-time tracking. The kernel method provides a general way of linearization and can be independent of the specific model with efficient computing. Introducing the kernel learning method into target tracking is expected to improve environmental adaptability. In this paper, based on the idea of kernel method, the current research progress of kernel learning target tracking is presented, including target detection method based on kernel learning, generative and discriminative target tracking method, and multi-kernel learning method with different kernel functions. Further research on kernel learning target tracking for kernel function optimization, long-term tracking, feature extraction and target occlusion are prospected.

Keywords

• |kernel learning method|nonlinear mapping|target detection|target tracking|multiple kernel learning|pattern recognition