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

Overview

• This study assesses the accuracy of a vision-based algorithm designed to estimate the groundline in videos of trotting horses, using multiple camera angles including handheld devices, as a step towards objective and portable equine gait analysis.

Introduction and Background

• Equine lameness diagnosis is traditionally performed through subjective visual assessments, which may be influenced by observer bias and variability.
• Objective tools currently include marker-based motion capture systems, force plates, and inertial measurement units (IMUs), but these require specialized equipment setups that are not always portable or practical in all settings.
• Vision-based algorithms present an attractive alternative because they can analyze video footage without markers, potentially offering a portable, easy-to-deploy solution for gait analysis.
• However, the reliability and accuracy of vision-based methods must be established, especially across different camera positions, angles, and in less controlled environments.

Objective of the Study

• To evaluate the performance of a custom-designed vision-based algorithm for estimating the groundline in video recordings of trotting horses.
• To test the algorithm using multiple camera setups, including static cameras positioned at various angles and a handheld device, simulating real-world use cases.
• To examine how handheld camera use affects variability and accuracy of the groundline and stride parameter estimations.

Methods

• Participants: Eight Standardbred trotter mares were chosen for treadmill locomotion recording.
• Experimental Setup:
  • Horses trotted on a high-speed treadmill.
  • Seven iPhones were positioned at varying heights and angles around the treadmill; one of these was handheld to assess movement impact.
• Data Capture:
  • Video frames were analyzed by a deep neural network algorithm trained to identify and place 2D keypoints on anatomical landmarks of interest.
  • The keypoints were the eye, withers, and croup (tuber sacrale).
• Groundline Estimation:
  • Two scenarios were compared: the algorithm-estimated groundline versus a fixed groundline based on treadmill reference.
• Signal Analysis:
  • Vertical Displacement Signals (VDS)—vertical movements relative to the groundline—were computed for each keypoint across strides.
  • Stride-level parameters assessed included maximum (Maxdiff) and minimum (Mindiff) vertical displacement values.
  • Bland-Altman analysis, scatter plots, and histograms were employed to quantify agreement and differences between estimation methods.
• Impact of Handheld Camera:
  • Variability and accuracy metrics for handheld camera data were directly compared to those from a static camera to test robustness under less controlled recording conditions.

Results

• The algorithm-estimated groundline closely matched the fixed treadmill reference groundline.
• Statistical accuracy included near-zero mean angle error and a low mean average error (MAE) of 0.45° over 242,192 frames analyzed.
• Stride-level maximum and minimum vertical displacement differences (Maxdiff and Mindiff) showed an MAE of only 0.5 mm across 36,981 strides.
• Handheld camera use introduced slightly higher variability, but relative errors remained clinically acceptable with a stride-level MAE below 1.8 mm across 357 trials.
• The handheld method’s performance indicates the algorithm is robust even with camera movement and less stable positioning.

Conclusions and Implications

• The vision-based algorithm reliably estimates the groundline and important stride parameters using diverse video input setups, including handheld recordings.
• This result suggests potential for portable, markerless, and less resource-intensive gait analysis tools in equine practice.
• Limitations include the controlled treadmill environment and that only one horse breed and coat color were tested, which may limit how broadly the findings apply.
• Future research should validate the algorithm in various real-world conditions, diverse horse populations, and compare results against other objective gait measurement technologies.