Automated Head Impact Detection in Youth Ice Hockey: A Two-Stage Deep Learning Approach

Abstract

Detecting head impact events in youth ice hockey is a critical problem at the intersection of computer vision, sports safety, and public health. Despite increasing awareness of the long-term neurological risks of repetitive head trauma, most youth leagues lack systematic monitoring. This thesis addresses that gap by developing and evaluating an automated detection system tailored to the resource constraints of community hockey. The contributions are threefold. First, we introduce the first publicly available annotated dataset for this task, featuring a hierarchical labeling scheme for both general and head impact events. Second, we design a hierarchical two-stage pipeline for rare-event discovery: Stage 1 detects general impact events to prune routine play, and Stage 2 applies specialized classification to determine head involvement. Third, we provide a comprehensive empirical comparison between a player-centric approach (Model A) and a full-frame multi-modal fusion approach (Model B). Experiments show that the optimal deployment strategy uses a player-centric architecture for Stage 1 and a full-frame architecture for Stage 2. The best Stage 1 model (Model A: TSM + Motion) achieves 75% recall and 25% precision (F1-score 0.375) for general impact detection, successfully reducing the portion of video requiring human review from 100% to just 6.3%. For the subsequent head impact classification task, the best Stage 2 model (Model B: RGB + Flow + Pose) achieves 80% recall and 67.6% precision (F1-score 0.733). This results in an end-to-end head impact event detection recall of 60%. Technical analyses reveal several key findings: motion (optical flow) is the dominant signal for impact detection; Temporal Shift Modules (TSM) consistently outperform Inflated 3D Convnets (I3D) in both accuracy and computational efficiency; and pose estimation features provide limited value for this task due to low resolution and equipment occlusion. This work demonstrates the feasibility of automated safety monitoring in youth hockey and provides a foundational dataset and methodology for future research.