基于可拆分结构的中层属性行为识别

Activity Recognition of Middle-Level Attribution Based on Detachable Structure

摘要: 随着行为识别研究的深入，复杂行为识别受到了越来越多研究者的关注。然而复杂行为中存在的大量冗余信息以及噪声严重降低了行为识别的准确性。针对这一问题，本文提出了基于可拆分结构的复杂行为中层属性时序一致性结构，并采用动态时间规整算法进行行为时序模式匹配。通过对视频行为时序进行拆分，学习行为的中层属性表达，构建行为中层属性的一致时序关系，有效去除了视频中的噪声和冗余信息，提高行为识别的准确率，并且能够从行为时序发展角度解释行为，增强了行为分析的可解释性。在Olympics数据集上的实验结果表明，该算法能够有效提高行为识别的准确率。

Abstract: With the deeper research of activity recognition, more researchers pay attention to complex activity recognition. However, a large number of redundant information and noise reduce the accuracy of activity recognition. In order to solve this problem, the paper proposes temporally consistent middle attributes structure and matching patterns of activity by dynamic time warping algorithm. Through splitting videos and learning middle attribute of segments, we build the consistent temporal structure of each activity category effectively solving noise and redundancy. And it explains the process of activity from the perspective of timing and enhances the interpretability of activity analysis. The experiment results on Olympics dataset show that the algorithm can improve the ac-curacy of activity recognition effectively.