Gait tracking in dogs using DeepLabCut: A markerless machine learning approach for controlled settings

Gill, Harry ORCID: https://orcid.org/0000-0003-4809-9479, Charles, James, Grant, Robyn A ORCID: https://orcid.org/0000-0002-3968-8370, Gardiner, James ORCID: https://orcid.org/0000-0003-1902-3416, Bates, Karl and Brassey, Charlotte ORCID: https://orcid.org/0000-0002-6552-541X (2025) Gait tracking in dogs using DeepLabCut: A markerless machine learning approach for controlled settings. Applied Animal Behaviour Science, 292. 106769. ISSN 0168-1591

Preview

Published Version Available under License Creative Commons Attribution. Download (2MB) | Preview

Abstract

Analysing locomotion is critical for assessing canine health and diagnosing musculoskeletal conditions, yet traditional motion capture methods for dog gait analysis remain impractical in many clinical and industry settings. Markerless deep-learning approaches, such as DeepLabCut (DLC), offer a promising alternative, but their performance in gait analysis, particularly across diverse dog breeds, remains largely untested. In addition, the ability to automate aspects of gait parameter extraction from the resulting dataset, an important requirement for industry practitioners, is also widely untested. In this study, we trained a bespoke neural network on a 2100 training frames, for 2D markerless tracking on eight dog breeds and developed a scripted workflow for semi-automated gait parameter extraction. We calculated several temporal and kinematic variables, including duty factor and joint ranges of motion, comparing values of a widely studied breed (Labrador Retrievers) to literature data. Our model’s performance aligned with previous DLC studies, performing strongly on well-defined landmarks (E.g. nose, eye, carpal, tarsal), whilst struggling with less morphologically discrete locations (E.g. shoulder, hip). ANOVA results from our mixed model revealed a significant effect of body part on tracking performance (p = 0.003), yet no significant effect of breed (p = 0.828) and a small interaction effect between breed and body part (p = 0.049). Our semi-automated workflow successfully extracted gait parameters across our study breeds, though performance was highly dependent on the quality of underlying tracking data. Duty factor and stifle range of motion measures from our labradors showed good overlap with literature values, yet the broader distribution in our data highlighted important limitations in cross-study comparisons. These results suggest that a markerless deep-learning approach could provide a viable alternative to traditional motion capture for canine gait analysis, offering potential applications for both clinical and industry settings.