Real-time analysis of indoor sports game situations through deep learning-based classification

Live indoor sports broadcasts require dynamic camera control in response to relevant game situations such as penalties or timeouts, a process that traditionally relies on human operators. This paper presents a solution for automatic real-time classification of game states in indoor invasion sports, with handball as the primary case study and basketball as a secondary validation scenario. Our approach utilizes raw video directly from cameras, enabling real-time analysis. The system continuously processes video frames, assigning each to one of seven classes: left/right attack, left/right counterattack, left/right penalty, and timeout. A synthetic representation of each frame is used to standardize the depiction of game dynamics. The proposed pipeline includes object detection with a fine-tuned You Only Look Once (YOLO) model to locate players, the ball, and referees; object tracking to compute velocity vectors; generation of a synthetic frame representing the current game state; and final classification using a custom Dense Convolutional Network (DenseNet). Using a dataset of 20 handball matches, the proposed system achieved a macro-averaged F1-score of 96.1%, with a per-image inference time below 4 milliseconds, evaluated on 118,129 images from matches unseen during training. The same pipeline was subsequently applied to basketball using only two matches, achieving an F1-score of 92.5% on 12,390 images, thereby illustrating the transferability of the proposed approach to other indoor invasion sports. The full pipeline operates in 34.04 milliseconds with GPU acceleration, processing over 25 frames per second.