摘要： 为在复杂环境中对目标进行长期的精确跟踪，提出一种基于在线学习和结构约束的目标检测和跟踪算法。采用改进的光流法对特定目标进行自适应跟踪，实时目标检测采用非层次结构在线学习随机蕨丛分类器。用基于结构约束的非监督学习法精确确定目标位置，以适应目标的形态变化。实验结果表明，该算法能够适应目标的基本形态变化，在目标出现尺寸变化、旋转、部分遮挡或短暂消失时都能稳定精确地跟踪目标。

关键词: 随机蕨丛, 结构约束, 光流法, 在线学习, 目标跟踪, 前向-后向误差

Abstract: In order to accurately track target during a long term in complex environment, this paper presents an algorithm combining detecting and tracking based on online learning and construction constraint. Improved optical flow method is used to adaptively track specific target. The nonhierarchical structure online learning random ferns classifier is used as real time target detecting method. In order to adapt target shape variation, unsupervised learning method based on construction constraint is used to accurately determine the target position. Experimental results show that this algorithm can adapt to the basic target shape variations and track the target steadily when the targets change in size, rotate, shelter partly or disappear in a short term.

Key words: random ferns brake, construction constraint, optical flow method, online learning, target tracking, forward-backward error

中图分类号: 

• TP311