中图法分类号：tp3914 文献标识码：a 文章编号：1006-8961(200 .doc

中图法分类号：TP391.4 文献标识码：A 文章编号：1006-8961(200 ) - - 论文引用格式： 中层时空特征的人体行为识别王泰青，王生进 1.清华大学 电子工程系清华信息科学与技术国家实验室北京 摘 要：目的：人体行为识别是计算机视觉领域的一个重要研究课题，具有广泛的应用前景。本文针对局部时空特征和全局时空特征在行为识别问题中的局限性，提出了一种新颖有效的行为中层时空特征。方法：该特征通过描述视频中时空兴趣点邻域内局部特征的结构化分布，增强时空兴趣点的行为鉴别能力，同时，避免对人体行为的全局描述，能够灵活地适应行为的类内变化。本文使用互信息度量中层时空特征与行为类别的相关性，将视频识别为与之具有最大互信息的行为类别。结果：实验结果表明，本文提出的中层时空特征在行为识别准确率上优于基于局部时空特征的方法和方法在KTH数据集和日常生活行为（ADL）数据集上分别达到了结论：提出的中层时空特征通过利用局部特征的时空分布信息增强了行为鉴别能力能够有效地识别多种复杂人体行为。 关键词：行为识别；时空兴趣点；中层时空特征；点互信息Human Action Recognition Using Mid-Level Spatial-Temporal Features Wang Taiqing1, Wang Shengjin1 1. Tsinghua National Laboratory for Information Science and Technology, State Key Laboratory of Intelligent Technology and Systems, School of Electronic Engineering, Tsinghua University, Beijing 100084, China Abstract: Objective: Human action recognition is an important research topic in the field of computer vision and has promising application potentials. By analyzing the limitations of both local and global spatial-temporal features, a novel and effective middle-level spatial-temporal feature is proposed for action recognition. Method: This feature encodes the structural distribution of local features in the neighborhood of the spatial-temporal interest point (STIP), thus improving the discriminative power of STIP and capable to model the flexible intra-action variations. Pointwise mutual information is introduced to measure the correlation between the mid-level feature and the action, and the video clip is classified as the action category that has the greatest mutual information with the mid-level features. Result: Experimental results validated the advantage of the proposed mid-level feature over the local-feature-based baseline methods and other published results. The mid-level feature achieved 96.3% and 98.0% recognition accuracies on the KTH and ADL action dataset, respectively. Conclusion: The proposed mid-level spatial-temporal feature enhances the discriminative power for actions