Multi-View Collision Detection in Multi-Sport Videos

Abstract

Physical impacts in sport present significant health risks, particularly for children with hemophilia, a blood coagulation problem in which even minor impacts may cause internal bleeding and potentially lead to long-term joint damage. Automatically detecting and monitoring these impacts using video recordings would allow for a more objective and continuous evaluation of physical exposure during sporting activities. This would help better adapt sporting recommendations to the pathology of the patient.

This work addresses the problem of action spotting in multi-view sports videos recorded from two synchronized cameras. While action spotting has been extensively studied in single-view settings, extending it to synchronized multi-camera inputs remains a largely unexplored problem in the literature. To this end, we introduce a new annotated dataset of 47 multi-view sports videos covering four disciplines and containing 3,180 annotations across three categories: player-player, player-ground and player-equipment.

Using this dataset, three approaches are developed and compared: (i) an adaptation of T-DEED, an action spotting model extended for multi-view inputs, (ii) a hybrid pipeline combining 2D pose estimation, monocular depth estimation and graph-based player interaction modeling and (iii) a fully deterministic rule-based method using geometric heuristics derived from body representations.

The results reveal that all three approaches face significant challenges, primarily due to the limited size and high visual diversity of the dataset, which prevents any of the approaches from learning a consistent and generalizable representation of physical contact. Nevertheless, the improvement observed when adding the auxiliary view in the hybrid pipeline suggests that multi-view information can be beneficial when combined with structured representations. These findings highlight both the difficulty of extending action spotting to multi-view settings and the need for larger and more visually consistent datasets to make such systems practically viable.