Find in Library

3D human pose estimation (HPE) has become increasingly important in baseball analytic, but there are several difficulties pertaining to pose estimation in real-world baseball pitching. First, in-the-wild baseball pitching lacks related 3D pose datasets and contains lots of joints occluded by other body parts. Second, baseball pitching contains dramatic velocity changes during arm acceleration phases. Due to these properties of pitching, it is difficult to use common filters to remove random noises while preserving high-frequency critical joint movements in pitching. To solve these problems, we propose joint-wise volumetric triangulation to reconstruct 3D human poses by utilizing the information of multi-view 2D joint heatmaps generated by 2D HPE methods. We also designed a baseball-customized filter system to remove noisy signal from pose movement while preserving the high-frequency pitching motion. Our proposed pose reconstruction scheme yields a 33.1 mm average position error and 0.35m/s (1.28 km/h) average velocity error on baseball pitching motion. Our work can be directly applied to estimate human poses either in indoor environment or real-world baseball field.