摘要
基于机器学习的思想并充分利用外观信息,提出一种在线选择纹理和形状特征的混合随机朴素贝叶斯视觉跟踪器。构造归一化空间金字塔,通过强度二值特征和金字塔梯度方向直方图二值特征,描述全局与局部区域的纹理和形状;并根据特征描述的二值性和多模性,设计并实现了在线混合朴素贝叶斯分类器。分类器预测类别后验概率生成信任图,跟踪器通过分析信任图实现目标跟踪,并利用极大似然估计和交叉验证实现外观学习和特征选择。选用基准测试集比较同类方法,从性能和复杂度两方面评估了跟踪器。实验结果表明跟踪器对光照变化,部分遮挡等情况具有一定的适应能力,且执行速度较快,存储空间较小。
Based on the idea of machine learning and the sufficient appearance, a mixture random Nave Bayes visual tracker with online texture and shape feature selection is proposed. The texture and shape of global and local region is described with binary feature of intensity and pyramid histogram of oriented gradients using normalized spatial pyramid. An online mixture of Nave Bayes classifier is designed and realized according to binary and multimodel description. The classifier predicts the class posterior probability to generate the confidence map, then the tracker analyzes the confidence map to track the object, learns the appearance with maximum likelihood estimation,and selects the feature with cross validation. Compared with homogeneous methods, the tracker is evaluated with performance and complexity based on benchmarks. The experimental results show that the tracker has certain adaption to illumination change and partial occlusion, and fast execution speed as well as little memory space.
出处
《光学学报》
EI
CAS
CSCD
北大核心
2015年第3期195-205,共11页
Acta Optica Sinica
基金
国家自然科学基金(11272347)
国家973计划(2013CB733100)
作者简介
郭鹏宇(1985-),男,博士研究生,主要从事视觉跟踪与运动分析等方面的研究。E-mail:pengyu.guo@nudt.edu.cn;通信联系人:张小虎,E-mail:zxhl302@hotmail.com;于起峰(1958-),男,教授,中国科学院院士,主要从事空天图像测量与视觉导航等方面的研究。E-mail:yuqifeng@vip.163.com
