Journal DOI 10.68956/journal.human-performance-futures

Creative Commons source reading · Motion capture and movement imaging

A Markerless Approach for Full-Body Biomechanics of Horses

Sarah K. Shaffer; Omar Medjaouri; Brian Swenson; Travis Eliason; Daniel P. Nicolella

Originally published in Animals (Basel) · 2025-08-05

From the article

Selected source text. JournalOS removes markup and shortens excerpts where indicated. The full article provides the complete context.

Abstract

Simple SummaryEquine gait data can be used to identify lameness and provide sports feedback. However, most tools that measure how a horse moves are time-consuming to use and require specialized expertise and equipment. This study introduces a method to track full-body, three-dimensional equine motion using videos collected by multiple cameras. Unlike traditional methods, no instrumentation needs to be placed on the horse. A neural network is used to identify the location of anatomic landmarks in the videos (e.g., the horse’s fetlock) and these predictions are tied to a musculoskeletal model. This process allows joint angles, stride information, and other motion data to be quickly computed. Our results show that most joint angles calculated by this method are within 10 degrees of those determined using data collected by a traditional…

3.2. Kinematics Results

For the PFERD_base network, median prediction error was lower for axial markers (4.1–4.5 cm) compared with appendicular markers (5.1–5.3 cm) regardless of gait (Figure 5; Table A1). However, the prediction error had high variance and were right skewed (Table A1, Figure 5), likely due to noise among marker predictions. Boxplots of prediction error for each marker are provided in Figure S2 of the Supplementary Materials. At the walk and trot, the poll (marker 1, C_Poll), withers marker (marker 10, C_Back_1), and right tuber coxae (marker 25, R_Pelvis_1) had an average prediction error greater than our exclusion threshold and were not used for IK. At the canter, the poll and withers markers were similarly excluded. For the PFERD-SwRI_Horse network, the median prediction error was similar for axial (3.2–4.1 cm) and appendicular marker…

4. Discussion

This study demonstrates that the proposed three-dimensional markerless motion capture pipeline is feasible in horses. Our base neural network (PFERD_base) predicted skeletal landmarks with good accuracy and our pipeline calculated most kinematic trajectories within 10° of the ground truth. To demonstrate that additional data would improve predictions, we also tested our pipeline using a network trained over data from more horses in a wider variety of environments (PFERD-SwRI_Horse). This dataset contained only slightly more images than the base network and were sparsely labeled but included data on more horses in more environments. Decreases in both the mean and variance of marker prediction error indicate that this network provided more accurate and less noisy predictions than the base network. In addition, it resulted in less error…

Editorial reading guide

  • How was movement measured, and what error or calibration information is reported?
  • Can the finding transfer from the tested task and population to the intended sport setting?