Reconstruction Accuracy Assessment of Surface and Underwater 3D Motion Analysis: A New Approach

Abstract

This study assessed accuracy of surface and underwater 3D reconstruction of a calibration volume with and without homography. A calibration volume (6000 × 2000 × 2500 mm) with 236 markers (64 above and 88 underwater control points—with 8 common points at water surface—and 92 validation points) was positioned on a 25 m swimming pool and recorded with two surface and four underwater cameras. Planar homography estimation for each calibration plane was computed to perform image rectification. Direct linear transformation algorithm for 3D reconstruction was applied, using 1600000 different combinations of 32 and 44 points out of the 64 and 88 control points for surface and underwater markers (resp.). Root Mean Square (RMS) error with homography of control and validations points was lower than without it for surface and underwater cameras (P≤0.03). With homography, RMS errors of control and validation points were similar between surface and underwater cameras (P≥0.47). Without homography, RMS error of control points was greater for underwater than surface cameras (P≤0.04) and the opposite was observed for validation points (P≤0.04). It is recommended that future studies using 3D reconstruction should include homography to improve swimming movement analysis accuracy.