Abstract: When the tracked objects get seriously obscured or have fast moving, the self-training method based on online semi-supervised boosting will lead to the error accumulation thus easily suffering from the drifting issue or even tracking failure. To overcome the disadvantages, a novel object tracking algorithm is proposed based on online semi-supervised boosting with positive（P） and negative（N） constraints, termed PN-SemiT. The proposed algorithm trains a binary classifier by using online semi-supervised boosting algorithm, and the training process is guided by structural constraints which restrict the labeling of the unlabeled set. Therefore, P-N constraints can evaluate the classifier, identify examples that have been classified in contradiction with structural constraints and adjust the?classifier?error?in?real-time. Experimental results on several different challenging video sequences show that the proposed algorithm has a superior tracking performance, and can alleviate the object drifting problem under complex environments.

Key words: object tracking, online learning, semi-supervised learning, object drifting, structural constraints

摘要： 针对在现有的基于在线半监督boosting的目标跟踪算法中，当目标发生遮挡或快速移动导致分类器更新过程中有错误引入时，其自训练机制会造成分类器错误累积进而产生跟踪漂移甚至导致跟踪失败的问题，提出了一种基于结合正负样本约束的在线半监督boosting的目标跟踪算法（简称PN-SemiT）。该算法在原有的在线半监督boosting跟踪算法的基础上，通过增加正负样本约束条件来实时纠正分类器的错误，并且将目标的先验模型和在线分类器相结合，通过不断迭代更新分类器来预测未标记样本的类别标记和权重。实验结果表明，与传统的在线半监督boosting目标跟踪算法和其他跟踪算法相比，PN-SemiT具有更优异的跟踪性能，能够在复杂的跟踪环境下有效缓解目标跟踪漂移问题。

关键词: 目标跟踪, 在线学习, 半监督学习, 目标漂移, 结构约束