Automated analysis and highlight generation for handball matches

Abstract

Manual analysis of handball matches is time-consuming and prone to human error, often failing to capture key events in fast-paced gameplay. This thesis presents a real-time, end-to-end system that automates the detection, tracking, and classification of player actions and generates highlight reels from match footage. The system integrates a YOLOv12-based object detection model, ByteTrack for multi-object tracking, and a Long-term Recurrent Convolutional Network (LRCN) combining CNN and LSTM layers for action recognition. Explainability is incorporated through Grad-CAM and LIME visualisations to enhance model transparency. The system was trained on the UNIRI-HBD dataset and evaluated on match-like scenes. The action recognition model achieved an F1-score of 84%, with a top-3 accuracy of 98.5%. The highlight generation module produced concise and relevant video segments, validated through expert review. While effective in recognising key handball actions such as jump shots and passes, the system showed limitations in distinguishing more nuanced behaviours like defence and dribbling. The methodology demonstrates promise for scalable applications in sports analytics and may be adapted for similar team sports with further domain-specific tuning.