ISSN 1000-1239 CN 11-1777/TP

• 论文 • 上一篇    

基于自适应多特征融合的mean shift目标跟踪

袁广林1,2 薛模根1,3 韩裕生3 周浦城3   

  1. 1(合肥工业大学计算机与信息学院 合肥 230009) 2(解放军炮兵学院二系 合肥 230031) 3(解放军炮兵学院四系 合肥 230031) (ygl6904@sina.com)
  • 出版日期: 2010-09-15

Mean Shift Object Tracking Based on Adaptive Multi-Features Fusion

Yuan Guanglin1, 2, Xue Mogen1, 3, Han Yusheng3, and Zhou Pucheng3   

  1. 1(School of Computer and Information, Hefei University of Technology, Hefei 230009) 2(Second Department, Artillery Academy of PLA, Hefei 230031) 3(Forth Departrment, Artillery Academy of PLA, Hefei 230031)
  • Online: 2010-09-15

摘要: 经典mean shift目标跟踪算法简单快速,具有较好的跟踪效果,但是它用单个特征描述目标,易受相似目标与背景的干扰,鲁棒性较差.针对此不足,推导出多特征融合mean shift目标定位公式;为了适应跟踪过程中目标与背景的变化,提出利用概率分布可分性判据动态评价特征对目标与背景的区分能力,并自适应地计算特征融合权重.在上述两个方面的基础上,对mean shift目标跟踪算法进行了改进,提出一种多特征融合mean shift目标跟踪算法.实验结果表明:提出的算法比经典mean shift目标跟踪算法具有更强的抗干扰性能和较高的跟踪精度.

关键词: 目标跟踪, 均值漂移, 特征融合, 核函数直方图, 边缘方向直方图

Abstract: Traditional mean shift tracking algorithm has achieved considerable success in object tracking due to its simplicity and robustness. But, it models the objects to be tracked with single color feature, which leads to that it is more prone to ambiguity, especially if the scene contains other objects characterized by a color distribution similar to that of the object of interest. To address this problem, a formula for target localization with mean shift based on multiple features is deduced. In addition, a method that evaluates the discriminability of each features with respect to foreground to background separability and adaptively calculates the features fusion weight by probability separability criterion is proposed. With the above deduced formula and proposed method, a novel mean shift target tracking algorithm based on adaptive multiple features fusion is presented. The proposed algorithm is run for each feature independently and the output of the mean shift algorithm for each feature is weighted based on the fusion weight. The states of the target in the current frame are computed through the integration of the outputs of mean shift. Experiments are conducted with color sequences and gray sequences, and its results show that the proposed algorithm has better performance than the classical mean shift tracking algorithm.

Key words: target tracking, mean shift, features fusion, kernel function histogram, edge orientation histogram