Automated facial analysis for working dog fatigue assessment

Working dogs perform safety-critical tasks, yet field fatigue assessment relies on subjective handler observation or invasive, impractical instrumentation. We present a non-invasive, video-only framework that extracts 46 anatomically grounded (DogFACS-based) facial landmarks from short recordings and produces fatigue assessment. Unlike wearable physiological monitoring, it attaches nothing to the animal; unlike manual DogFACS coding, it is automated and scalable; and unlike raw-video deep learning, it uses a compact, interpretable representation robust to coat color and lighting and suited to small field datasets. Landmark trajectories were classified by a nested model-selection pipeline spanning tree-based models and deep sequence models. On 177 videos from four working dogs recorded under field conditions, the pipeline reached mean F1-scores of 0.781 for tree-based models and 0.789 for neural-network models in dog-specific nested cross-validation, and 0.676 and 0.740, respectively, under nested leave-one-dog-out validation. We also introduce a dog-specific Relative Facial-State Index, giving strong held-out before–after discrimination in three dogs, and a mouth-motion analysis whose aperture-magnitude features discriminated within all four dogs. To our knowledge, this is the first automated facial-landmark approach to working-dog fatigue assessment under operational conditions, motivating validation in larger, more diverse cohorts.