Research Note: Deep Learning Model for Gait Analysis in Dogs

Source

Pahk J-H, Park S-J, Seo J-H, et al. A deep learning–based markerless gait analysis model for dogs shows promising accuracy when validated with 2-dimensional marker–based data. Am J Vet Res. 2026:1-8. doi:10.2460/ajvr.25.09.0337

Research Note

Subjective visual assessment of gait analysis is prone to observer bias and variability among evaluators. Quantitative gait evaluation using methods like force plate analysis and infrared camera–based optical motion-capture systems offer reproducible and data-driven measures for collecting information but require expensive equipment and complex infrastructure. In human studies, markerless gait analysis techniques can provide data accuracy comparable to marker-based approaches while decreasing preparation time and costs.¹

This study^a trained and validated a deep learning–based markerless gait analysis model for use in dogs. Client-owned dogs (n = 408) representing >30 breeds and a variety of body types were recorded walking or trotting back and forth along a 3.3-m straight walkway with a smartphone camera positioned laterally at the midpoint of the walkway. Recordings were captured at 30 frames per second; 20,000 markerless images from 374 dogs were used for model training. A marker-based dataset using physical markers placed on key anatomic landmarks on the thoracic and pelvic limbs was collected for model validation and testing using 3,566 and 568 images from 20 and 14 dogs, respectively.

Results demonstrated high accuracy of the markerless model for capturing canine anatomic landmark positions (mean average precision, 96.6%), although performance varied by anatomic region, with distal landmarks showing higher accuracy than proximal landmarks.

Study results suggested this system may be useful for clinical canine gait assessment; however, further validation across breeds and environments is needed.

^a This study was supported by AIFORPET Corp.