Abstract
Abstract
With the rapid growth of the scale of film and television content and the increasing demand for intelligent editing processes, problems such as large time scale spans, strong spatial interference, and weak continuity of action semantics in human actions in film and television works have become increasingly prominent. Therefore, the research aims to construct a human action recognition framework that integrates multi-scale visual Transformer and spatial–temporal attention mechanisms. Multi-scale context encoding is used to describe the long-term and short-term temporal correlation of actions, and multi-scale spatial and temporal attention networks are introduced to explicitly enhance key human body regions and key action fragments. Cross-scale feature fusion and action aggregation are combined to achieve robust discrimination. When the occlusion ratio reached 30%, the model recognition performance remained above 0.86, and the fluctuation amplitude was controlled within 0.02. In the transitional action detection task, the key action hit rate reached the highest 0.96, showing the ability to effectively capture the action switching process. The average detection delay was stable at around 30 ms, which was about 20% lower than that of the comparison method. The proposed multi-scale modeling and attention collaboration strategy can effectively improve the accuracy, stability and robustness of human action recognition in complex film and television scenes, providing a feasible technical path for intelligent film and television editing and content understanding.