141 Locomotion pattern analysis using digital video labeling by machine learning

The assessment of lameness has long relied on the expertise of a human observer. Innovations in technology have provided new options to objectively detect lameness, but most systems involve equipment attached to the horse, potentially causing changes in their normal locomotion patterns. One such tool, the Lameness Locator, analyzes motion data recorded by accelerometers mounted to the horse with bandages, stickers or hats. Tactile sensors in horse skin across the body can detect pressures as small as 0.008g, which is comparable to the sensitivity of a human fingertip. Thus, the weight and instability of even very small accelerometers may cause variation in the way horses move when being tested for lameness. In this study, we used a software package to digitally label videos of individual horses with and without the Lameness Locator applied. Unlike current methods of detection, this software can be used in the natural environment without attaching equipment to the animal. Thus, we hypothesize that the digital video analysis approach will objectively identify shifts in locomotion patterns due to the pressure from the Lameness Locator device. All horses were video recorded at 120fps and a resolution of 1280 × 720p on a standard path while handled in halter at the walk, trot and canter for both the left and right-side views. The DeepLabCut software package was used to extract coordinates for 23 skeletal landmarks as they proceed across the frame. Quantitative gait parameters such as stride length, stance time, and vertical travel of the withers were derived from the landmark tracking data by a robust data processing pipeline. The pipeline comprises a series of operations, including outlier removal due to occlusion, trajectory smoothing, model-fitting-based phase identification, and gait parameters calculation. In this preliminary analysis we considered the stride length, stance time, and the range of travel for the poll, withers and croup at the trot; key metrics traditionally used to assess lameness. Comparing these gait variables within each horse, we found that the right-to-left normalized stride length, stance time, and normalized withers range were all significantly different for each horse between the unencumbered and Lameness Locator equipped trot passes (P