Athlete 3D MOTION from video: application to injury prevention in on-field and off-field environments

‍1. Introduction

Injury is an inevitable consequence of most professional sports, especially contact sports. Injuries arise from unsafe movements, on and off pitch. Multi-camera marker (eg Vicon) and markerless (eg SIMI) systems are the gold standard for athlete kinematic tracking to understand injury prevention, but these are available only for certain elite athletes. Tracking athlete kinematics using one or more handheld cameras holds promise for much more widespread use. This paper presents approaches to tracking athlete collisions on-pitch using three calibrated mobile cameras as well as tracking rehabilitation motion using a single camera.


‍

‍

2. Materials and Methods

On-pitch scenario: we staged low-severity rugby ruck collisions recorded with three spatially calibrated cameras. The 2D human joint positions were inferred using open-source software [1]. Camera calibration (MATLAB tool) combined with using algebraic triangulation allowed for these 2D poses to be lifted to 3D space, centred on the pelvis position.Off-pitch scenarios, we used eight volunteer athletes to perform common strength & rehabilitation exercises (squats, counter movement jumps (CMJs) and Romanian deadlifts (RDLs)) in a Vicon motion capture environment (eight infra red cameras), where we also recorded the athlete movements using various mobile devices (mobile phones and tablets) in a range of positions in an arc around the athlete. Detectron2 was applied for 2D key point detection [2] together with Stridedformer [3] to predict 3D pose. Knee joint and torso angle changes were assessed by comparison with the Vicon outputs.


3. Results

Sample on-field kinematic results are shown in Figure 1. Sample error measures are shown for the rehabilitation exercises in Table 1 for a camera at 20 degrees from the side and 3m distance from the athletes.

‍

‍

4. Discussion and Conclusions

The on-pitch tracking using three calibrated cameras is successful up to the point of occlusion, and this approach could be used together with Kinepose [4] to initialise multibody dynamics simulations to estimate collision forces from video [5].The accuracy of the single camera tracking of common rehabilitation exercises shows significant promise as a means to quantify, though further evaluation is needed.

5. References

1. Cao et al., IEEE Mach Learning, 20192. Lo & Girshick, Facebook 20193. Li et al, CVPR, 2022.4. Gildea et al, CVPR 20225. Reid et al IRCOBI 2022.

Acknowledgements: Enterprise Ireland & Science Foundation Ireland