Abstract: video-based human action recognition has received much attention recently in computer vision while it is a very challenging research with many applications such as automatic behavior analysis, visual surveillance and human-computer interaction. A human action recognition method based on cascaded structure is proposed in this paper. Firstly, a trajectory-based method is proposed to select the interest points detected by the Dollar detector, which is sensitive to image noise, camera movement and zooming. Therefore, the pseudo interest points in the background can be effectively excluded and the extracted features will be more relevant to action recognition. Secondly, an automatic feature selection method based on the combination of normalized cuts and mRMR criteria is used to determine a subset of the words generated by the Bag-of-Words model and construct a cascaded structure for action recognition. The purpose is to make the feature used by the cascaded structure more distinctive. Lastly, the promising experimental results validate our contribution to the improvement of accuracy in human action recognition.

Key words: space-time interest point selection, normalized cuts, mRMR criteria

摘要： 基于视频的人体动作识别是近年来计算机视觉领域备受关注且十分具有挑战性的研究方向,可以应用于人的行为分析，视频监控和人机交互等方面。本文提出了一种基于级联结构的人体动作识别方法:针对Dollar时空兴趣点检测器易受图像噪声、摄像机运动与缩放等因素影响产生伪兴趣点的问题，提出了一种基于轨迹差异度的兴趣点筛选方法，有效避免了引入背景中的伪兴趣点，提高了人体运动特征提取的准确度；采用规范切与mRMR准则对词袋模型生成的特征向量进行自动特征选择，同时建立一个用于分类的级联结构，在识别各类不同动作时选择不同的特征子集，使得分类器使用的特征更具区分性。在KTH人体运动测试集上实验，验证了文中方法能提高动作识别的准确度。

关键词: 时空兴趣点筛选, 规范切, mRMR准则