视频群体行为识别综述

群体行为识别是指给定一个包含多人场景的视频,模型需要识别出视频中多个人物正在共同完成的群体行为.群体行为识别是视频理解中的一个重要问题,可以被应用在运动比赛视频分析、监控视频识别、社交行为理解等现实场景中.多人场景视频较为复杂,时间和空间上的信息十分丰富,对模型提取关键信息的能力要求更高.模型只有高效地建模场景中的层次化关系,并为人物群体提取有区分性的时空特征,才能准确地识别出群体行为.由于其广泛的应用需求,群体行为识别问题受到了研究人员的广泛关注.对近几年来群体行为识别问题上的大量研究工作进行了深入分析,总结出了群体行为识别研究所面临的主要挑战,系统地归纳出了6种类型的群体行为识别方法,包含传统非深度学习识别方法以及基于深度学习技术的识别方法,并对未来研究的可能方向进行了展望.

Group Activity Recognition in Videos: A Survey

Given a video containing a multi-person scene, group activity recognition model needs to recognize the group activity that multiple people in video are completing together. Group activity recognition is an important problem in video understanding and can be applied to sports videos analysis, surveillance video recognition, social behavior understanding, and other real scenarios. Multi-person scene video is complicated, and the spatial-temporal information is rich, which requires the model to extract key information. To accurately recognize group activity, the model should efficiently model the hierarchical relationships in the scene and extract distinguishing spatial-temporal features for people. Due to its wide range of application requirements, the problem of group activity recognition has received extensive attention from researchers. This study has conducted an in-depth analysis of a large number of research work on group activity recognition in recent years, and summarized the main challenges of group activity recognition research, systematically summarized six types of group activity recognition methods, including traditional non-deep learning recognition methods and recognition methods based on deep learning technology, and proposed the possible directions of future research.