Pose Estimation for Evaluating Standing Long Jumps via Dynamic Bayesian Networks

A system is developed for analyzing poses in a standing long jump automatically. In the system, silhouette of the jumper is segmented from the background first. A thinning algorithm is then used to find a rough skeleton from the silhouette. Some image processing techniques are applied to make the resulted skeleton smoother and simpler. Key points are extracted from the skeleton. Finally, the dynamic Bayesian network (DBN) is used to determine the corresponding pose from the key points. The experimental result shows that pose estimation accuracy is quite good. According to the standing long jump standards, incorrect movements at different stages of the jump can thus be identified.