Automated Detection of Physical Contact Events in Youth Ice Hockey: A Player-Focused Deep Learning Approach

Abstract Youth ice hockey lacks scalable, automated tools to quantify collision exposure and flag head-impact events from routine game video. This study presents a player-centric, video-based pipeline that detects physical contacts and uses them as a high-frequency proxy to accelerate head-impact dataset creation and support injury surveillance. Broadcast footage was segmented into 30-frame clips. Players were detected using a youth-tuned You Only Look Once, version 8 (YOLOv8) model and tracked with StrongSORT augmented by an intersection over union (IoU)–based cost term. Contact events from 20 games (U11–U15) were manually annotated to generate contact-centered clips; non-contact clips were sampled to create datasets with contact: non-contact ratios of 1:4, 1:8, and 1:12. A Temporal Shift Module (TSM) classifier operated on player crops, with crop scale and class balance systematically evaluated. Performance was assessed using precision, recall, F1, specificity, Matthews correlation coefficient (MCC), and area under the precision–recall curve (PRAUC), and compared with a full-frame baseline. The youth-specific YOLOv8 achieved precision 0.96, recall 0.97, average precision (AP)0.7 0.94, mean average precision(mAP)50 0.97, and mAP50–95 0.82. IoU-assisted tracking reduced identity (ID) switches by 69% (172→53) and improved Multiple Object Tracking Accuracy (MOTA) from 89.0% to 94.5%. The best classifier (15 segments, shift-division = 4, 1.5× crop, 1:12 ratio) achieved precision 0.96, recall 0.94, F1 0.95, specificity ≈1.00, and MCC 0.95. On two full-game tests, PRAUC was 0.51; at a 0.5 threshold, precision was 0.25 and recall was 0.79. As a head-impact pre-filter, the system detected 19/22 impacts(86.4% recall) and reduced review time from over three hours to under 30 minutes per game. This pipeline combines youth-specific detection, IoU-enhanced tracking, and temporal context–aware classification to deliver a scalable, efficient solution for collision exposure monitoring and dataset generation, supporting safer, data-driven youth hockey environments.