ARCHIVES

This study investigates video-based handball action classification using the UNIRI Handball Dataset Version 1. The study focuses on seven handball actions: crossing, defence, dribbling, jump-shot, passing, running, and shot. The dataset was organized by associating scene videos, labelled action clips, and available player-detection information using scene identifiers. To reduce the possibility of information leakage, the data were divided at the scene level so that action clips originating from the same scene remained within a single subset. The study evaluates a sequence of feature representations and learning approaches, beginning with MediaPipe Pose-based pose and motion descriptors and conventional machine-learning classifiers, followed by temporal deep-learning models using recurrent, temporal convolutional, skeleton-based, and Transformer-based architectures. Based on these experiments, a video-based spatial-temporal architecture was developed using a pretrained ResNet18 network for frame-level spatial feature extraction, followed by feature projection, a bidirectional gated recurrent unit (BiGRU) for temporal modelling, and a temporal attention mechanism for aggregating informative features across the video sequence. A fixed number of RGB frames was sampled from each action clip to construct the temporal representation. The ResNet18 backbone was used as a frozen feature extractor, while the temporal and classification layers were trained using class-weighted loss with label smoothing. Model selection was performed using validation data, while the test subset was kept locked until final evaluation. The proposed framework provides a structured and leakage-aware approach for learning spatial and temporal representations from handball video and addresses the challenge of distinguishing visually and temporally similar handball actions.