时间连续贝叶斯分类目标跟踪算法

Temporal consistent Bayes classification object tracking algorithm

为提高目标跟踪的鲁棒性,提出一种基于结构稀疏表示的时间连续贝叶斯分类跟踪算法。在粒子滤波框架下进行,采用结构稀疏表示原理对样本进行线形重构。考虑到跟踪过程中目标形态帧间的连续性,将时间连续约束项嵌入线性重构目标方程,设计目标方程求解方法,获得稀疏系数;为更好地提取稀疏系数中的相似度信息,利用贝叶斯原理设计一款分类器,通过跟踪过程中获得的正负样本进行训练,有效地对候选目标进行分类。将该算法与其它4种先进的算法在6组测试视频中进行比较,实验结果表明,该算法在复杂条件下具有较高的鲁棒性。

For improving the robustness of object tracker,a structured sparse representation based temporal consistent Bayes classification tracking algorithm was proposed.Under the framework of particle filter,the structured sparse representation principle was used to linearly recombine samples.To encourage the consecutiveness of inter-frame,the temporal consistency constraint term was imbedded into the objective function.The coding coefficients were obtained by designing a solving method of the function.For better extracting the likelihood information from coding coefficients,a classifier based on the principle of Bayes was designed.The classifier was trained by both positive and negative samples,and the candidates were classified effectively.The proposed tracker was compared with 4state-of-the-art trackers on 6testing videos.Experimental results demonstrate that the proposed tracker is more robust in complex scenes.