Ball Trajectory and Spin Analysis From Asynchronous Videos

Existing systems for ball trajectory and spin estimation use embedded sensors or expensive high-frame-rate cameras, which severely limits their accessibility. We propose an easy-to-setup low-cost vision sensor pipeline using two static asynchronous consumer-grade cameras. We also propose the use of epipolar geometry for synchronizing the cameras. We estimate 3-D ball trajectory and spin with only one distinguishable feature on the ball. Mixture of Gaussians and adaptive color-based thresholding are used to localize the ball in 2-D followed by triangulation. To estimate spin magnitude and axis, we employ feature detection and plane fitting. Extensive experiments with three different balls across multiple varied environments are reported and the approach is validated by arriving at the standard gravitational acceleration value from our estimated ball trajectory. For validating the spin, we compare our results with the true spin for a rotating ball fixed on a motor shaft. The average reprojection error was below 10 pixels for all our experiments and a maximum deviation of 17 rotations per minute in spin magnitude was observed.