计算机工程与应用 ›› 2016, Vol. 52 ›› Issue (6): 172-177.

• 图形图像处理 • 上一篇    下一篇

基于混合观测模型的粒子滤波跟踪算法

吴  桐1,2,王  玲2,李钟敏1,何  凡1   

  1. 1.中国洛阳电子装备试验中心,河南 洛阳 471000
    2.湖南大学 电气与信息工程学院,长沙 410082
  • 出版日期:2016-03-15 发布日期:2016-03-17

Particle filter tracking algorithm based on mixture appearance models

WU Tong1,2, WANG Ling2, LI Zhongmin1, HE Fan1   

  1. 1.Luoyang Electronic Equipment Examination Center of China, Luoyang, Henan 471000, China
    2.College of Electrical and Information Engineering, Hunan University, Changsha 410082, China
  • Online:2016-03-15 Published:2016-03-17

摘要: 为了提高目标外观迅速变化时视觉跟踪算法的鲁棒性,提出了一种基于混合观测模型的粒子滤波跟踪算法。在粒子滤波构架下,使用加权核直方图模型结合mean shift算法对粒子进行初定位,通过正交子空间模型作为精确的观测模型,估计目标的最终状态。这样既能迅速地学习到目标外观变化的趋势,又避免了使用正交子空间而产生的跟踪漂移。实验结果表明,该算法在光照变化、姿态变化、遮挡的情况下,均具有较强的鲁棒性。

关键词: 视觉跟踪, 子空间, 自适应学习, 粒子滤波, mean shift

Abstract: In order to improve the robustness of visual tracking algorithm when the target appearance is rapidly changing, this paper presents a particle filter tracking algorithm based on mixture appearance models. On the particle filter framework, the histogram is utilized to roughly localize the object by a mean shift procedure. A more precise eigenspace appearance model is invoked to infer the final state of the object. In this way, it will not only learn the changing trends of the target appearance rapidly, but also avoids the tracking drift by using the orthogonal subspaces. The experimental results show that the algorithm can keep strong robustness in the light change, posture change, occlusion.

Key words: visual tracking, subspace, adaptive learning, particle filter, mean shift