基于局部分块学习的在线视觉跟踪

Online Visual Tracking Based on Local Patch Learning

视觉跟踪中,如何构建一种能够适应目标表观特征变化的目标模型是增强算法跟踪精度和稳定性的关键之一.本文提出利用跟踪区域内像素的初始分类标记来构建目标的局部分块模型,并在贝叶斯理论框架下提出了基于局部分块学习的在线视觉跟踪算法.首先,利用标定的初始跟踪区域构建目标的局部分块模型;然后,在当前跟踪区域中通过局部分块学习和贝叶斯估计确定当前帧的跟踪结果;最后,利用特征聚类对局部分块模型进行更新.实验结果表明:所提算法对目标表观变化的适应性明显增强,跟踪精度和稳定性较近年来的同类算法均有一定提高.

In visual tracking,how to construct an object model to cope w ith the appearance change is one of the key problems to improve tracking precision and stability. To resolve this problem,this paper proposes to construct a local patch model using the initial labels of the pixels in tracking area,and proposes an online visual tracking algorithm based on local patch learning under the framew ork of Bayesian theory. The detailed operation is as follow s. Firstly,it constructs the local patch model according to the initialized tracking area. Then,it utilizes the object model to learn the local patches in current tracking area and estimates the current state via Bayes estimation. Finally,it updates the local patch model by feature clustering. The experiment results indicate that the proposed algorithm obtains a distinct improvement in coping w ith appearance change,and exceeds the recent local patch-based trackers in both tracking precision and stability.