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

Abstract

Background: Equine lameness diagnosis largely relies on subjective visual assessments, which can be biased. Although marker-based methods, force plates and inertial measurement units (IMUs) provide objective measurements, they require specialized setups. Vision-based algorithms offer a portable, markerless alternative, but their accuracy needs thorough testing.

Objectives: To evaluate a custom vision-based algorithm for estimating the groundline across multiple camera angles, including handheld use in horses trotting on a treadmill.

Study design: Experimental comparative study.

Methods: Eight Standardbred trotter mares were recorded trotting on a high-speed treadmill using seven iPhones positioned at various heights and angles, including a handheld device. A trained deep neural network algorithm placed 2D keypoints on each video frame. Vertical Displacement Signals (VDS) for the eye, withers and croup (tuber sacrale) were computed relative to either an algorithm-estimated or a fixed treadmill groundline. Maximum (Maxdiff) and minimum (Mindiff) stride values were compared using Bland-Altman analysis, scatter plots and histograms. The effect of handheld use on variability and accuracy was assessed by comparing results from a handheld camera to those from a static camera.

Results: Groundline estimation closely matched the fixed reference, exhibiting near-zero mean angle error and low mean average error (MAE = 0.45°; n = 242.192). Maxdiff and Mindiff stride-level (n = 36.981) MAE were 0.5 mm, with clinically acceptable additional variability introduced by handheld use at the trial level (Maxdiff and Mindiff MAE < 1.8 mm; n = 357).

Main limitations: Treadmill-based data and a single breed/coat colour may limit generalizability to other settings.

Conclusions: The vision-based algorithm accurately estimates the groundline and stride VDS parameters from various camera setups, including handheld. Further validation in diverse environments and against other objective gait analysis systems is recommended.

Keywords: deep learning; equine lameness; groundline estimation; handheld; lameness detection; objective gait analysis; pose estimation; vision‐based algorithm.

FIGURE 1

Camera positions and angles relative to the treadmill and horse. The dynamic estimated groundline is visible near the horse's hooves. The animations in the corners show the camera angle as seen from above.

FIGURE 2

Experimental setup. The study population (n = 8) were recorded from six angles as described in Figure 1 with iPhones recording at HD, 30 FPS. The frames were keypoint‐annotated by a trained deep neural network. The groundline was either fixed or dynamically estimated by using the hoof keypoints, and the VDS were filtered. The resulting VDS signals were stride split, and the matching strides were compared statistically by comparing Mindiff and Maxdiff. The VDS curves on the right illustrate matching strides and a Maxdiff calculation as an example.

FIGURE 3

The vertical displacements (red, green, blue arrows) are measured in the 2D sagittal plane of the horse (blue square), when seen from the side, as the distance in pixels from the dynamically estimated groundline (orange) or a fixed groundline to the keypoints. The pixel difference is calibrated to a metric value, knowing the metric withers height.

FIGURE 4

Histograms of signed groundline angle error (left) and absolute groundline angle error (right).

FIGURE 5

Scatter plot (top), Bland‐Altman plot (middle) and histograms of signed and absolute differences (bottom) for MinDiff values comparing estimated versus fixed groundlines across all keypoints.

FIGURE 6

Scatter plot (top), Bland‐Altman plot (middle) and histograms of signed and absolute differences (bottom) for MaxDiff values comparing estimated versus fixed groundlines across all keypoints. Bland–Altman outliers outside plot: (19.6, −44.1; −18.8, 21.6).

FIGURE 7

All keypoints—Stride‐based comparison between handheld and stationary camera. Scatter plot (top), Bland‐Altman plot (middle) and histograms of signed and absolute differences (bottom) for Maxdiff and Mindiff values comparing handheld versus stationary cameras across all keypoints (n = 6229).

FIGURE 8

All keypoints—Trial‐based comparison between handheld and stationary camera. Scatter plot (left), Bland‐Altman plot (right) and histograms of signed and absolute differences (bottom) for Maxdiff and Mindiff values comparing handheld versus stationary cameras across all keypoints (n = 357).