A review of video action recognition algorithm based on deep learning

Abstract

As a core research branch within computer vision, video-based human activity classification techniques bring significant improvements to intelligent surveillance, medical health monitoring, and human-machine interactive systems in terms of safety guarantee, operational efficiency and user experience. Early human motion analysis methods rely solely on static single-frame images, which fail to satisfy practical application demands amid the rapid development of intelligent vision technologies. Modern research communities mainly focus on motion recognition tasks based continuous video sequences. This survey systematically sorts out designed for action classification. We first elaborate the research background and typical structural designs of prevailing deep learning models, together with a detailed comparison of their respective strengths and limitations. Then, standard benchmark datasets, quantitative evaluation metrics and performance comparison schemes widely adopted in this field are summarized. Finally, existing bottlenecks restricting the practical deployment of video motion recognition systems are discussed comprehensively.