视频序列中的行为识别研究进展

Research progress on activity recognition in video

视频序列中的行为识别是计算机视觉领域非常活跃的研究方向,在视频监控、人机交互、医疗看护、视频检索等领域有着广阔的应用前景。综合分析了行为识别的研究进展,首先分析了目前主流的行为识别数据库的不同特点;然后总结了行为的特征表示方法以及识别方法的最新进展,由于实际应用要求算法具有快速、自动和实时性,且人工标定大量视频存在困难,无监督的学习方法受到越来越多的关注;在此基础上,讨论了行为识别面临的挑战,以及今后的研究发展方向。

Human activity recognition in video sequences is a very active research field of computer vision with broad application prospects in visual surveillance,human-computer interaction,medical care,video retrieval and etc.In this paper a comprehensive analysis of recent developments of activity recognition is given.Firstly,the different characteristics of current mainstream databases are analysed.Then the methods of the feature representation and the latest progress of the recognition algorithms are summarized.Especially,as the real applications require algorithm capable of fast,atomatic and real-time processing,and it is difficult to label a lot of videos manualy,unsupervised learning earns more and more attention.Finally,some detailed discussions on research challenges and future directions in activity recognition are also provided.