Mean Shift & Kernel Tracking 的介绍
The goal of kernel-based tracking is to match a fixed kernel-weighted histogram to a time and parameter varying one.
A Great Tutorial Mean Shift Tutorial
Kernel Tracking paper Kernel-Based Object Tracking
Kernel Tracking 的问题
- Unable to deal with multi-modal pdfs
- Fail in tracking small and fast moving targets
- Can not recover a track after a total occlusion
- Some motions cannot be recovered using kernel-based methods. such as rotational motion (with rotationally-symmetric kernels)
- Not scale-invatiant ( entended in paper Mean-shift Blob Tracking through Scale Space )
- Performance are largely effected by the quality of searching directions calculated from the measurements, i.e. the measurements become more or less invariant to some motion parameters, such that these motion parameters are not recoverable or observable
- Though mean-shift tracker performs well on sequences with relatively small object displacement, its performance is not guaranteed for objects in highly cluttered environment especially when it undergoes partial occlusion.
- As pointed out by Yang and Duraiswami, the computational complexity of traditional Mean-Shift Tracking is quadratic in the number of samples, making real-time performance difficult. Although the complexity can be made linear with the application of a recently proposed fast Gauss transform, tracking in real-time remains a problem when large or multiple objects are involved. We propose to boost the efficiency of mean-shift tracking using random sampling.
- The classical information-theoretic similarity measures such as the Bhattacharyya coefficient or the Kullback-Leibler divergence are not very discriminative, especially in higher dimensions.
Kernel Tracking 的优点
- Combine summary descriptions of both intensity values and spatial positions in a way that avoids the need for complex models of object shape, appearance or motion
- Low computational cost
Kernel Tracking 的介绍
最新推荐文章于 2021-04-28 21:24:35 发布