Validation of a handheld smartphone markerless gait-analysis tool using an estimated groundline in horses

A handheld smartphone-based computer vision algorithm (RealHorse® [RH]) offers accessible alternatives for equine gait analysis but requires validation against a gold-standard three-dimensional multicamera optical motion capture system (Qualisys® [QS]).

To evaluate the accuracy and precision of RH in measuring vertical displacement signals (VDS) at the eye, withers, back and croup in horses trotting on a straight line and on a circle.

Cross-sectional comparative validation study of a markerless computer vision algorithm.

Fifty-nine horses were recorded while trotting on a straight line and 24 were lunged on a circle. RH detected two-dimensional anatomical keypoints on each frame, which were used to estimate a dynamic groundline and compute ground relative VDS with stride-based difference in maxima (Maxdiff) and minima (Mindiff). QS provided synchronous ground-relative VDS reference values. Agreement was evaluated using mean signed error, mean absolute error and Bland-Altman analysis.

On the straight line (n = 2620 strides), the pooled stride-level MAE for Maxdiff and Mindiff was 3.8 mm. Keypoint-specific errors were 5.1 mm (eye), 4.3 mm (withers) and 3.0 mm (croup). On the circle (n = 2419 strides), pooled stride-level error increased to 5.5 mm. Trial-level analysis (n = 58 trials) showed much lower errors: 1.4 mm for both eye and withers and 1.1 mm for croup. On the circle (n = 24 trials), trial-level errors were higher, with 2.8 mm for the eye, 1.8 mm for the withers and 3.3 mm for the croup. The back keypoint consistently showed the lowest errors across both stride and trial levels.

RH measurements of the croup Mindiff during circling resulted in higher values and showed the largest error.

RH measured vertical displacement of all keypoints with high accuracy and precision (trial-level MAE 1.1-1.4 mm straight, 1.8-3.3 mm circle), supporting its use for equine gait analysis.