Reliability, agreement and variability of a markerless computer vision algorithm for equine gait analysis under field conditions.
Authors: Key Karsten, Berg Katja, Kirkegaard Jakob, Andresen Kristian Ringkjær, Skov Hansen Sabrina
Journal: Equine veterinary journal
Summary
# Editorial Summary Markerless computer vision technology promises to democratise equine gait analysis by eliminating the need for reflective markers or specialised equipment, yet its accuracy under real-world field conditions has remained largely unvalidated. This cross-sectional study assessed a proprietary iPhone-based algorithm's ability to detect vertical displacement signals (VDS) at three anatomical landmarks—eye, withers and croup—across 67 recordings from 37 horses trotting on straight lines and circles, comparing algorithmic keypoint detection against manual annotation by independent observers. At the stride level (1556 strides analysed), mean absolute errors were impressively consistent: the eye achieved 2.9–3.0 mm accuracy, the croup 4.3–4.4 mm, and the withers 5.5 mm for both maximal and minimal vertical displacements, with trial-level analysis showing even tighter agreement (2.3–3.7 mm across landmarks). Subjective lameness scoring correlated reasonably well with these objective measurements, though the authors acknowledge that validation against established gait analysis systems and further refinement of groundline estimation remain necessary steps before widespread clinical adoption. For practitioners seeking accessible, field-deployable gait assessment tools, these findings suggest the algorithm is sufficiently reliable for detecting meaningful changes in vertical motion dynamics, particularly at the eye where precision is greatest—a significant development for remote monitoring or initial lameness screening in practice settings where traditional motion capture systems are impractical.
Read the full abstract on PubMed
Practical Takeaways
- •Markerless computer vision offers a practical field-based gait analysis tool for equine practitioners, achieving sub-centimetre accuracy for detecting vertical displacement changes during trotting
- •Eye and croup measurements are more reliable than withers measurements with this algorithm; eye keypoint is most sensitive for detecting subtle gait changes (2.9 mm accuracy)
- •While algorithm shows promise for objective lameness assessment, direct clinical comparison with established gait analysis systems is still needed before replacing traditional diagnostic methods
Key Findings
- •Frame-level vertical keypoint accuracy was 4.5 mm (eye), 5.5 mm (croup), and 11.8 mm (withers) using markerless computer vision algorithm
- •Stride-level mean absolute errors for vertical displacement maxima and minima were 4.3 mm overall, with eye keypoint showing lowest errors (2.9-3.0 mm) and withers highest (5.5 mm)
- •Trial-level analysis across 67 recordings demonstrated consistent algorithm performance with lower absolute differences (2.3-3.7 mm), indicating reliability across multiple strides
- •Algorithm robustly measured vertical displacements under varied field conditions using handheld iPhone recordings from straight lines and circular paths