Reliability, agreement and variability of a markerless computer vision algorithm for equine gait analysis under field conditions.

Overview

• This study evaluates the reliability and accuracy of a markerless, computer vision-based algorithm designed to analyze horse gait by measuring vertical displacement signals at specific anatomical points during trotting under field conditions.
• The research compares algorithm outputs with manual annotations to assess its precision and consistency in practical use.

Objective and Background

• Equine gait analysis traditionally requires specialized equipment and markers on horses, which can be invasive and less accessible in field conditions.
• Computer vision-based markerless algorithms offer a non-invasive, accessible alternative but need thorough validation for reliability and accuracy under varied conditions, such as different movement patterns (straight lines, circles) and environments.
• This study specifically assesses a proprietary vision-based algorithm focusing on vertical displacement signals (VDS) at the horse’s eye, withers, and croup, and evaluates groundline estimation essential for accurate movement measurement.

Methods

• Data Collection:
  • 67 handheld iPhone video recordings of 37 horses trotting on straight and circular paths were collected in field settings.
• Algorithm Processing:
  • Both the vision-based algorithm and independent manual annotation identified 2D anatomical keypoints on every video frame.
  • From these keypoints, the algorithm estimated the groundline—used as a reference baseline for vertical measurements—and computed vertical displacements (VDS) along with stride-based maxima (Maxdiff) and minima (Mindiff) vertical differences.
• Comparative Analysis:
  • Mean signed error (MSE), mean absolute error (MAE), and Bland-Altman plots were used to compare keypoint locations and VDS metrics between algorithm outputs and manual annotations.

Results

• Keypoint Detection Accuracy:
  • At the frame level, the average vertical keypoint errors were:
    • Eye: 4.5 mm error
    • Croup: 5.5 mm error
    • Withers: 11.8 mm error
  • Manual annotation error was lower at approximately 2.7 mm, representing a baseline human-level precision.
• Stride-Level Vertical Displacement Accuracy:
  • Analysis over 1556 strides showed mean absolute errors for both Maxdiff and Mindiff of 4.3 mm overall.
  • By anatomical point:
    • Eye had the smallest errors (Maxdiff: 2.9 mm, Mindiff: 3.0 mm), indicating the highest precision here.
    • Withers exhibited the highest errors (5.5 mm for both Maxdiff and Mindiff), possibly related to movement variability or detection challenges.
    • Croup errors were moderate (Maxdiff: 4.3 mm, Mindiff: 4.4 mm).
• Trial-Level Analysis (Per Recording):
  • Across 67 trials (recordings), signal stability improved when aggregating multiple strides, reflected by lower absolute differences:
    • Eye: 2.3 mm
    • Withers: 3.7 mm
    • Croup: 2.7 mm
  • Despite fewer strides per trial than optimal, the algorithm maintained consistent performance.
• Subjective Validation:
  • Comparisons with subjective lameness scoring showed alignment with objective gait metrics, though some variation was noted, indicating potential clinical relevance and areas for further refinement.

Conclusions and Implications

• The computer vision algorithm robustly measured vertical displacement signals of horses’ eyes, withers, and croup under diverse field conditions, both straight and circular trotting.
• Groundline estimation, critical for accurate VDS measurement, was validated separately in treadmill data, stressing the algorithm’s adaptability and precision.
• Errors were generally low, close to manual annotation precision, demonstrating the algorithm’s reliability for stride-level and trial-level gait analysis.
• There is potential for clinical application in equine movement assessment, especially given the alignment with subjective lameness evaluations.
• The study recommends further comparisons with established gait analysis systems to solidify the algorithm’s clinical utility and extend validation across more settings and horse populations.
• Overall, this technology could enable accessible, non-invasive, and accurate gait assessments for veterinarians and researchers using everyday devices like smartphones, facilitating field-based equine health and performance monitoring.