Evaluation of a custom trained human pose estimation model for kinematic analysis of pole vault performance

Kinematic outputs from a custom human pose estimation (HPE) system and a marker-based system (MB) were compared. Six trained/highly trained pole vaulters (two males and four females) participated in a single testing session of 2-8 vaults. HPE utilized footage from three 50 hz cameras and a pole vault-trained model (based on the ASPset21j dataset) for tracking. Vaults were temporally normalized from take-off to peak pelvis height. Centre of mass (COM) and joint centre (JC) locations were compared between systems using Bland-Altman analysis, Mean Absolute Error (MAE) in all three planes, linear regression (R2), and average Euclidean distances. Peak COM heights ranged from 2.98 to 3.94 m across participants. COM position demonstrated an adequate level of agreement in pole vault context (Bland-Altman bias: 9-63 mm, MAE: 37-71 mm, R2 : .95-1.00, average Euclidean distance: 99 ± 41 mm). Between-system comparisons varied across different JCs (Bland-Altman bias: 3-93 mm, MAE: 30-99 mm, R2: 0.40-0.99, Euclidean distance: 85-127 mm). Importantly, the vertical COM MAE of 41 mm is less than the minimum increase in bar height typically applied in pole vault competition (50 mm), indicating that the HPE measure of COM height is sufficiently accurate for evaluating vault performance.