Abstract: Target tracking in complex scenes has been a challenging issue facing the research communities worldwide. In this paper, a random local mean-hash feature based tracking method is proposed. The core of the proposed approach lies with the target modeling using Poisson probability distribution and model updating by online adaptive learning. The presented algorithm is put to test against other four existing learning-based standard tracking algorithms, that is, OnlineBoostingTracker, SemiTracker, BeyondSemiTracker, Context Tracker and MILTracker. The results indicate that the presented algorithm achieves better comprehensive tracking performances, particularly in occlusion resistance. The improvement is needed in the algorithm’s anti-rotation performance which will be the focus in the feature study.

Key words: mean-Hash, multiple instance learning, online learning, object detection and tracking, classifier

摘要： 在局部遮挡，光线变化，以及复杂背景环境下进行有效稳定的目标跟踪一直是一个长期困扰研究者的复杂问题。提出一种基于随机局部均值Hash特征的在线学习目标跟踪算法，算法的创新点为基于泊松概率分布的目标模型建立及其在线更新。算法首先利用已标定实际位置的目标图像来初始化目标模型及构建初始分类器池，由此求出下一帧的检测算子，同时基于多实例在线学习方法，利用检测到的目标样本（正样本）以及附近的背景样本（负样本）在线更新目标模型，求出新的检测算子用于后续帧的目标检测及跟踪。提出的算法与现有基于检测学习的OnlineBoostingTracker，SemiTracker，BeyondSemiTracker，Context Tracker和MILTracker跟踪算法在给定的四个标准视频序列中进行了跟踪性能比较。实验结果表明，在各种复杂环境下，该算法具备良好的综合跟踪性能，尤其在抗局部遮挡方面尤为突出。在抗目标旋转方面，该算法仍有待优化。

关键词: 均值Hash, 多实例学习, 在线学习, 目标检测与跟踪, 分类器