Computer Vision and Deep Learning during Sport Collisions

Injuries are a major problem in sport. We aim to understand the forces applied to the human body during injury events in sport, and how this canbe prevented based on biomechanical factors such as speed, player orientation and technique. Concussion in rugby will be used as a test case in the project, though the methods can be applied to any sports injury, such as ankle sprain and ACL rupture.

The principal objective is to combine head kinematic data (available from wearable sensors) and player whole-body configurations (extracted from video using computer vision approaches) with computational modelling (forward dynamics multibody and finite element models) to understand how player impact forces depend on player configurations at the instant of injury in sport. A unique wearable sensor dataset (instrumented mouthguards) involving elite men’s and women’s rugby teams will be available. Multi-camera view video footage from these games will be used to drive state-of-the-art human pose estimation algorithms for tackler and ball-carrier whole-body configurations before, during and after collision events. We will map the relationship between the measured sensor data and the player speed and poses for concussive and non-injury control cases. These human pose estimates will also be used to configure computational collision models and permit the evaluation of “what if” scenarios to address the likely effects of altered player posture and speed on subsequent collision force and injury probability, thus informing injury prevention strategies.