Real-Time Action Detection in Volleyball Matches Using DETR Architecture

Mu-Jan Shih, Yi-Yu Hsu

Abstract: Real-time action detection in sports, particularly in dynamic and fast-paced games like volleyball, poses substantial challenges for traditional object detection models. Models like YOLO (You Only Look Once), which rely heavily on single-frame analysis, often fall short in capturing the rapid and intricate movements characteristic of such sports. This paper introduces a novel real-time action detection model based on the Detection with Transformers (DETR) architecture, leveraging the self-attention mechanism to effectively process global information across consecutive frames. The elimination of non-maximum suppression (NMS) significantly reduces inference time variability, enhancing its real-time capabilities. Our model supports detailed individual action detection (identifying specific actions performed by single players) and robust group activity predictions (recognizing interactions among multiple players), which are particularly useful for team sports analytics. Experimental results indicate that our model delivers high accuracy and immediacy in capturing and analyzing players’ actions. Improved action detection accuracy significantly enhances strategic analysis and training, enables precise automated video editing for highlight generation, and enriches the spectator experience by providing more engaging and informative content.