Evaluating the Accuracy of a Vision-Based Algorithm for Groundline Estimation in Trotting Horses Using Multiple...

Area of Science:

• Veterinary Medicine
• Biomechanics
• Computer Vision

Background:

• Equine lameness diagnosis often relies on subjective visual assessments.
• Objective methods like IMUs require specialized equipment.
• Vision-based algorithms present a portable, markerless alternative needing validation.

Purpose of the Study:

• To evaluate a custom vision-based algorithm for groundline estimation.
• To assess accuracy across multiple camera angles, including handheld use.
• To analyze its application in horses trotting on a treadmill.

Main Methods:

• An experimental comparative study involving eight Standardbred mares.
• Used iPhones for video recording at various angles, including handheld.
• Applied a deep neural network to estimate 2D keypoints and compute Vertical Displacement Signals (VDS).

Main Results:

• Groundline estimation showed near-zero mean angle error and low mean average error (MAE = 0.45°).
• Stride-level MAE for Maxdiff and Mindiff was 0.5 mm.
• Handheld use introduced clinically acceptable additional variability (Maxdiff and Mindiff MAE < 1.8 mm).

Conclusions:

• The vision-based algorithm accurately estimates groundline and VDS parameters from diverse camera setups.
• Limitations include treadmill-based data and single breed/coat color, potentially affecting generalizability.
• Further validation in varied environments and against other objective systems is recommended.