基于贝叶斯模型与最佳伙伴相似度量的目标跟踪  

Visual object tracking using Bayesian models and best-buddies similarity

作　　者：徐福来 王鸿鹏[1,2] 张普 赵仲奇 刘景泰 Xu Fulai;Wang Hongpeng;Zhang Pu;Zhao Zhongqi;Liu Jingtai(Institute of Robotics ＆ Automatic Information System,Nankai University,Tianjin 300350,China;Tianjin Key Laboratory of Intelligent Robotics,Tianjin 300350,China;Baihe Nature Reserve,Aba Sichuan 623400,China)

机构地区：[1]南开大学机器人与信息自动化研究所,天津300350 [2]天津市智能机器人技术重点实验室,天津300350 [3]四川省阿坝州白河自然保护处,四川阿坝623400

出　　处：《计算机应用研究》2018年第8期2498-2502,2526,共6页Application Research of Computers

基　　金：国家自然科学基金资助项目(61375087);天津市自然科学基金重点项目(15JCZDJC31200);国家留学基金资助项目

摘　　要：提出了一种基于产生式与判别式联合模型的视觉目标跟踪算法。首先介绍了一种基于全局颜色特征直方图特征的贝叶斯分类器,检测出若干最有可能属于目标的候选区域;然后利用最佳伙伴相似性度量(best-buddies similarity)得到候选区域与目标模板的相似度,结合概率值与相似度值估计出最优的目标状态。通过划分目标—背景区域模型、目标—干扰区域模型,对可能产生干扰的区域提前进行抑制,降低了长期跟踪可能产生的漂移问题的风险;同时引入了自适应尺度估计机制和在线模型更新策略,以获得更为精准的跟踪结果。在37组具有挑战性的图像序列上与七种优秀的算法对比实验表明,所提出的算法能够有效应对光照变化、遮挡、旋转与尺度变化等多种问题。This paper presented an object tracking algorithm using both the generative appearance model and the discriminative model. First, the algorithm employed a Bayesian classifier based on holistic color histogram to distinguish the object. To avoid the tracker drifting to nearby regions which appear similarly to the object of interest, it identified and suppressed the distracting regions in advance. Second, the algorithm adopted a generative model based on best-buddies similarity（BBS） to measure the simi-larity between candidate regions with the object template. Finally, the algorithm introduced a multiple scale searching method to enable the tracker to handle scale change, and used an adaptive model update strategy to enhance the tracking performance. Extensive experimental results on 37 challenging sequences show that the proposed algorithm performs favorably against state-of-the-art methods in terms of accuracy and robustness.

关 键 词：目标跟踪 贝叶斯分类器 最佳伙伴相似度量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]