基于最大互信息区域跟踪的人体行为检测算法

Human Action Detection Based on Tracking Region of Maximum Mutual Information

人体行为检测问题不仅需要判断行为的类别,而且需要估计行为发生的时间和位置,有重要的现实应用意义.人体行为检测的主要难点在于参数空间维度高以及背景运动干扰.针对上述难点,本文提出了一种基于最大互信息区域跟踪的人体行为检测算法.该算法将行为区域定义为最大互信息矩形区域,采用稠密轨迹作为底层特征,利用随机森林学习轨迹特征与行为类别的互信息函数,利用轨迹的时间连续性对行为区域进行大时间跨度的预测和跟踪.实验结果表明,该算法不仅能够有效地识别不同类别的行为,而且能够适应现实场景中背景运动的干扰,从而准确地检测和跟踪行为区域.

Human action detection tries to estimate not only the category but also the time and place of the action, which are significant for real-world applications. The main difficulties of action detection lie in the high dimensionality of the parameter space and the distraction of dynamic background. To attack these difficulties, an algorithm based on tracking of the maximum-mutual-information region is presented in this paper. We define the action region as the region of maximum mutual information. We use dense trajectories as the low-level feature, learn the mutual information function between the trajectory and the action category via the random forest. Then, the action region is tracked in a large-time span by making use of the continuity of trajectories. The experimental results show the effectiveness of our method in recognizing different actions, and the ability of accurately tracking the action region in large-time-span in the presence of dynamic background.