Pose Estimation and Machine Learning for Sprinter Form Comparisons

Abstract

Our study aims to discover what joints and other factors play the most important role in developing faster sprinters. Using data from high schoolers and professional athletes, differences were compared in each athlete's top speed position at their MVP stance, which is the maximal height of vertical projection as defined by the ALTIS Kinogram Method. With TensorFlow's MoveNet Thunder pose estimation model, each athlete's leg and arm limb angles were investigated. Alongside these angle key points, the ground-to-air ratios (number of frames with foot on the ground divided by number of frames with foot in the air) of each athlete at their top speed were measured. Comparing the angles, it was discovered that professional athletes tend to have a more constricted range of movement at their maximal height projection. In comparison, school athletes tended to swing wider in the same position. With speed data of the high school athletes, it was also found that a shorter ground-to-air ratio also meant a typically faster time. Paired with this data, models also determined that the front hip and elbows determine athletes' speed. However, while these findings may serve as benchmarks for success, performance will vary among athletes, so individualized training will most benefit a sprinter.