基于深度学习和层次化运动建模的行为识别算法

The Action Recognition Algorithm Based on Deep Learning and Hierarchical Motion Modeling

随着计算机视觉和互联网技术的迅速发展,行为识别技术在智能视频分析和人机交互领域中的应用越来越广泛。该领域的一个核心问题就是如何建立一个高效的运动模型来捕捉视频中的运动信息。针对该问题,提出了一种基于自监督方式的分层对比运动学习框架,用于从原始视频帧中提取有效的运动表示。具体来说,该方法通过逐步学习网络中不同级别的层次化运动特征,从而减小了低层运动信息和高层识别任务之间的语义鸿沟,促进了多层次外观和运动之间的信息融合。我们提出的运动学习模块具有轻量化、灵活度高等特点,方便嵌入到现有的各种深度网络。从4个主流动作识别数据库上的大量实验结果表明,所提出的方法通过逐渐地捕捉更高级别的运动特征,并进化成对动作识别更有区分性的语义动态,相比于其他主流的行为识别方法可以获得更加优异的识别性能。

With the rapid development of computer vision and Internet technology,action recognition technology is widely used in the field of intelligent video analysis and human-computer interaction.A key problem in this field is how to build an efficient motion model to capture the motion information in video.To solve this problem,this paper proposes a hierarchical contrast motion learning framework based on self supervision,which is used to extract effective motion representation from the original video frame.Specifically,this method reduces the semantic gap between the low-level motion information and the high-level recognition task by gradually learning the hierarchical motion features of different levels in the network,and promotes the fusion of appearance and motion information between multiple levels.The proposed motion learning module is lightweight with high flexibility,which can be easily embedded into various existing deep learning networks.A large number of experimental results on four mainstream action recognition benchmarks show that the proposed method can achieve better recognition performance than other state-of-the-art methods by gradually capturing higher-level motion features and evolving into more discriminative semantic dynamics for action recognition.