Developing an Automated System to Measure Player Velocity for Head Impacts Evaluation using 2D Video in Youth Ice Hockey Games

Abstract

Traumatic brain injury (TBI) is a major concern in contact and collision sports such as football, rugby, ice hockey, and boxing, where athletes are frequently subjected to repetitive head impacts. These impacts range from mild concussions to severe neurological trauma and are often associated with long-term cognitive, behavioural, psychological, and physical impairments. In ice hockey, the combination of high-speed skating and physical contact significantly increases the risk of concussion. Concussion rates as high as 1.81 per 1000 athletic exposures have been reported, emphasizing the need for improved injury monitoring and prevention, particularly at the youth level, which constitutes the majority of participants in the sport. Young athletes are especially vulnerable due to their developmental stage and lack of access to monitoring technologies commonly available in professional settings. Many head impacts in youth hockey go undetected or unreported, impeding early intervention and increasing the risk of long-term effects. Effective injury mitigation requires a deeper understanding of the biomechanical factors contributing to brain trauma, including impact magnitude, frequency, inter-impact intervals, and cumulative exposure duration. Among these, velocity at the moment of impact plays an important role in determining the kinematics and energy transferred to the brain. However, the relationship between specific impact characteristics and clinical outcomes remains poorly understood. Developing predictive models requires large, structured datasets that capture biomechanical and contextual variables in real-world game settings. To address this need, and as part of a larger project aimed at developing a fully automated pipeline to extract head-impact characteristics from 2D video, this study focused on a primary step required for estimating head-impact velocity. Specifically, we developed and validated a fully automated pipeline for estimating planar player velocity using single-camera, side-view 2D video footage, an accessible and scalable solution tailored to youth hockey environments. The pipeline eliminates the need for specialized tracking systems and consists of three core components, structured into three studies: Study 1: Object Detection for Ice Surface Localization in Youth Hockey A YOLOv5 object detector was trained on 9,900 annotated side-view frames from youth games at the Peewee, Midget, and Atom levels to identify specific rink landmarks. When at least four landmarks were detected by the network, a custom script computed a homography matrix to project the side-view frame into a standardized top-down view. This approach demonstrated strong performance, achieving an F1 score of 0.99 and mAP@0.5 of 98.5% for object detection, along with an average Intersection over Union (IOU) of 0.96 for localization accuracy. Study 2: Development of an Automated System to Obtain the Planar Velocity of Individual Players From 2D Video of Youth Ice Hockey Games A YOLOv8 model, pretrained on 80,000 NHL images and fine-tuned on 4,700 youth hockey frames, detects and classifies players by jersey color (light/dark). Player tracking is performed using a modified StrongSORT algorithm, which reduced identity switch errors from 172 to 53 on 100 clips and improved Multi-Object Tracking Accuracy (MOTA) from 89.0% to 94.5%. Trajectories are transformed into top-down coordinates using the output of Study 1, and velocity is estimated based on displacement over 5-frame intervals (200 ms at 25 FPS), enabling real-time analysis of motion patterns relevant to head impact risk. A YOLOv8 model, pretrained on 80,000 annotated NHL player images and fine-tuned on 4,700 annotated youth hockey frames, was used to detect and classify players by jersey color (light/dark). The model achieved high detection accuracy, with a precision of 96% and a recall of 97%. Player tracking was performed using the StrongSORT algorithm enhanced with our custom cost function, which reduced identity switch errors from 172 to 53 across 100 clips (90 frames each at 30 FPS) and improved Multi-Object Tracking Accuracy (MOTA) from 89.0% to 94.5%. Player trajectories from side-view footage were transformed to a top-down view using frame-by-frame homography calculated in Study 1, and velocities were estimated based on displacement over 5-frame intervals (200 ms at 25 FPS). Study 3: Validation of Deep Learning-Based Velocity Estimation from Single, Side-View Ice Hockey Game Footage Using Top-Down Drone Measurements The objective of Study 3 was to validate the player velocities estimated by the proposed pipeline in Study 2 under real-world conditions, using drone-mounted top-down video as the ground truth. Velocities in the top-down footage were manually extracted in Kinovea by tracking a helmet-mounted marker across synchronized 10-frame intervals. Side-view estimates from both fixed and panning camera configurations were compared against the ground truth across four rink regions: two defensive faceoff circles and two central zones (near and far from the camera). Movement types included lateral, forward, and diagonal skating paths. A zoomed-in side camera configuration was also tested in the far zone to assess the impact of reduced landmark visibility. Results showed that both fixed and panning camera setups achieved mean absolute errors under 10%, while the zoomed-in configuration remained within 20%, a range still acceptable for categorizing impact severity. Bland–Altman analysis indicated strong agreement between the fixed/panning configurations and the top-down reference, providing preliminary validation of the pipeline. Sensitivity analyses were conducted by perturbing player coordinates in both axes by ±5, ±10, and ±15 pixels to evaluate susceptibility to localization error. The results revealed greater sensitivity in regions farther from the camera. Additionally, velocity was recalculated using temporal windows of 5, 10, 20, and 30 frames, showing that increasing the frame window consistently reduced error, particularly in the vertical direction where perspective distortion is more pronounced. This study proposed a preliminarily validated pipeline offering a cost-efficient, scalable, and accessible solution for estimating player velocity from standard broadcast footage of youth ice hockey games. Its fully automated structure and reliance on commonly available video make it particularly suited for large-scale injury surveillance and biomechanics research in environments without advanced tracking infrastructure or specialized sensor-based systems. As part of a larger project to develop an automated data capture system for documenting head impact characteristics, this study focused on creating and validating an AI-driven method for estimating player velocity in the context of head impacts. Velocity is a key factor influencing the energy transferred to the head during collisions. By generating large datasets and applying advanced AI algorithms, the system enables the identification of patterns linking impact characteristics to injury outcomes.