Fine-Grained Action Understanding in Instructional Sports Videos via a Hierarchical Spatiotemporal Pyramid Network
Songjiao Wu, Yuan Wang, Liping Wang · Traitement du signal · 2025
With the digital transformation of sports education and athletic training, the automated analysis and understanding of instructional sports videos have emerged as critical areas of research.Fine-grained action understanding models play an increasingly significant role in this context, as they are designed to accurately extract and analyze detailed motion information.Traditional approaches to action recognition have primarily relied on singlescale feature extraction, which has proven inadequate for handling complex spatiotemporal information, especially in scenarios characterized by high variability and rapid motion transitions.These limitations often result in reduced accuracy and poor real-time performance.In recent years, multi-scale network models have been explored to enhance video analysis capabilities; however, challenges remain in balancing computational efficiency with precision.To address these shortcomings, a fine-grained action understanding model based on a hierarchical spatiotemporal pyramid network was proposed in this study.By constructing a multi-scale spatiotemporal pyramid prediction algorithm, this model can improve the extraction of spatiotemporal feature points of sports actions.In addition, by incorporating a temporal scale-based fine-grained action prediction algorithm, the model can capture intricate details within instructional sports videos accurately.By optimizing dynamic spatiotemporal characteristics and temporal dependencies, this study achieves improved accuracy and real-time performance in the prediction of fine-grained sports actions, offering a novel theoretical and technical foundation for the development of intelligent sports instruction systems.