A service dog that limps under load at mile two of a work shift is not a reliable medical device. That dog is a liability to its handler, a future washout statistic, and in the worst cases, a dog whose pain went undetected for months because no one had the tools to see it. Markerless gait analysis from standard video footage is changing that. This article examines the current state of the technology, the model architectures that matter, and what the gap between human observation and machine detection actually looks like in practice.
Why Gait Is the Most Under-Evaluated Variable in Candidate Selection
Most service dog evaluation protocols focus heavily on obedience, task performance, environmental stability, and handler bond. These are the right things to measure. What the field has historically underweighted is musculoskeletal readiness under sustained working conditions.
A dog can pass a public access test while carrying a compensatory gait pattern that will produce chronic joint damage within eighteen months of full-time work. The compensation is often invisible at a casual trot on level flooring. It shows up in stride asymmetry data, in limb loading ratios during turns, in the progressive shortening of push-off phase across a thirty-minute evaluation session.
The International Association of Assistance Dog Partners has long advocated for health screening as a prerequisite to placement. Orthopedic clearances through OFA and PennHIP address structural concerns at a point in time. They do not capture functional gait under working conditions. That gap is where video-based markerless analysis operates.
At ServiceDog.AI, our position is that gait assessment should be a continuous variable tracked across a candidate's entire training career, not a one-time veterinary snapshot. The technology now exists to make that practical.
The Limits of Human Observation During Live Evaluation
Experienced trainers are remarkably good at detecting obvious lameness. A grade 3 or 4 lameness on the AAEP scale is visible to any competent observer. The clinical problem sits at grades 1 and 2, where lameness is inconsistent, subtle, or only apparent under specific loading conditions.
Human visual processing has documented limitations in detecting asymmetries below roughly 15 to 20 percent stride difference. Research in equine veterinary science has repeatedly shown that trained equine practitioners disagree substantially on lameness scores for horses in the grade 1 to 2 range. There is no reason to believe canine evaluation is materially different.
The evaluation environment compounds this. A trainer watching a dog for task compliance, environmental reaction, and handler cues simultaneously cannot allocate full visual attention to stride mechanics. The dog moving at a natural working pace covers ground faster than the human eye can resolve individual limb phase timing.
Video slows this down. Frame-by-frame review helps. But manual frame analysis at 30fps across a 20-minute evaluation session is not scalable. Automated pose estimation running on that same footage is.
How a Markerless Gait Analysis Pipeline Actually Works
Markerless tracking means no reflective markers, no force plates, no instrumented treadmills. The input is standard video from one or more cameras. The output is a time-series of skeletal keypoint positions from which gait metrics are derived.
The pipeline has four functional stages.
Detection: A bounding box detector localizes the dog in each frame. YOLOv8 and its successors perform well here, though any detector with strong small-object performance on natural environments works. Multi-dog scenes require instance tracking, which adds complexity but is solved by standard approaches like ByteTrack.
Pose estimation: A keypoint model maps anatomical landmarks onto the detected animal. Canine-specific keypoint vocabularies typically include nose, ears, shoulders, elbows, wrists, hips, stifles, hocks, and paw contact points. The number of keypoints varies by model but 17 to 24 landmarks is typical for biomechanical analysis. Models trained on general animal pose datasets (like Animal Kingdom or AP-10K) provide a starting point but require fine-tuning on canine-specific data for reliable stifle and hock tracking.
Temporal smoothing: Raw per-frame keypoint outputs are noisy. Kalman filtering, Savitzky-Golay smoothing, or learned temporal models (TCNs or transformers operating on keypoint sequences) reduce jitter without destroying the high-frequency signal components that encode heel strike and toe-off events.
Metric extraction: From smoothed keypoint trajectories, the system computes stride length, stride frequency, stance phase duration, swing phase duration, limb loading symmetry indices, step length asymmetry, and progressive fatigue indicators like stride shortening rate over time.
Model Selection: Canine Keypoints, Temporal Consistency, and Architecture Tradeoffs
The choice of pose backbone matters substantially for service dog evaluation contexts. Here is what the current landscape looks like and where the tradeoffs fall.
HRNet and ViTPose for Per-Frame Accuracy
HRNet maintains high-resolution representations through the entire network rather than downsampling and upsampling, which produces high per-frame keypoint accuracy. ViTPose applies vision transformer architectures to pose estimation and has shown strong cross-domain generalization. Both are reasonable choices for per-frame accuracy on canine subjects, but neither was trained on canine data and performance on hind limb keypoints in particular degrades when the dog is moving away from camera or at oblique angles.
Fine-tuning on a canine keypoint dataset annotated specifically for gait evaluation is not optional. Generic animal pose models produce unacceptable error rates on hock and stifle tracking, which are the joints most diagnostically relevant for the conditions you are screening for.
Temporal Models for Gait Cycle Consistency
Per-frame models are frame-independent by design. They do not know that the left hindlimb was in swing phase two frames ago. This produces phase inconsistency errors where the model flips keypoint labels or produces physiologically impossible joint angles on individual frames.
Temporal convolutional networks operating on keypoint sequences address this by learning the expected temporal structure of canine gait cycles. Video transformer approaches like those in the action recognition literature (adapted for regression rather than classification) can model longer gait cycle dependencies. For a trotting dog at 3 to 4 m/s, a full stride cycle is roughly 0.7 to 0.9 seconds, which at 30fps is 21 to 27 frames. A temporal model with a receptive field covering at least two full stride cycles provides the context needed for stable phase assignment.
Edge Deployment Constraints
Full ViTPose-Large inference at evaluation-time accuracy requires GPU resources that are not present in a phone or a lightweight evaluation tablet. Knowledge distillation from large teacher models to smaller student architectures (MobileNetV3, EfficientNet-Lite backbones) reduces this significantly. In our testing environment at ServiceDog.AI, distilled canine pose models running on a recent mid-range Android device with NPU acceleration achieve 28 to 30fps inference latency, which is sufficient for real-time gait scoring during live evaluation walks.
Clinical Signals ML Catches That Operators Miss
This is where the practical argument for automated gait analysis becomes concrete. These are the specific patterns that surface consistently in video analysis but escape live observation.
Compensatory asymmetry in diagonal pairs: Dogs with low-grade hindlimb pain frequently shift load to the contralateral forelimb. The forelimb compensation shows up as increased vertical displacement (head nodding) and extended forelimb stance phase. A trainer watching the dog from the side may notice nothing. Stride symmetry indices computed across the full session flag the pattern reliably.
Progressive fatigue-driven stride shortening: A healthy dog at working pace maintains relatively consistent stride length across a 20-minute session. A dog with early-stage hip dysplasia, unresolved soft tissue injury, or inadequate conditioning shows measurable stride shortening over time as fatigue removes the muscle support that was masking the underlying structural limitation. This rate-of-change metric does not exist in snapshot veterinary assessment.
Stance phase asymmetry during turns: Circular path walking loads the inside limb differently from straight-line movement. Dogs with unilateral stifle pathology often show marked stance phase asymmetry specifically during turning. A straight-line evaluation pass misses this entirely. A camera positioned to capture circular path movement, combined with automated phase detection, reveals it consistently.
Hock hyperextension under load: Certain heritable conditions produce subtle hock laxity that is not apparent at rest or slow movement. At a working trot under pack weight or repeated task repetition, hock extension angles exceed the normal range briefly at heel strike. The event lasts one to two frames at 30fps. Human observation during a live walk misses it nearly every time. Automated keypoint tracking flags it for review.
Deployment Realities: Edge Inference, Lighting, and Ground Surface Variation
Laboratory-grade gait analysis uses controlled lighting, calibrated cameras at fixed positions, and standardized surface materials. Service dog evaluation happens in training facilities, public access environments, parking lots, and handler homes. The deployment gap is real and requires explicit engineering decisions.
Lighting variation: Consumer cameras operating in auto-exposure mode produce frame brightness fluctuations that degrade keypoint confidence scores inconsistently. Normalizing input frames with histogram equalization or using models trained with aggressive photometric augmentation reduces this sensitivity. Evaluation sessions in direct sunlight with hard shadows on the dog's coat are the hardest case and remain an active area of model improvement.
Camera angle and calibration: Sagittal plane (side-on) video is the gold standard for stride length and limb phase analysis. Cranial-caudal (front/rear) views are necessary for medial-lateral sway and limb adduction/abduction metrics. A two-camera setup with known relative geometry enables 3D reconstruction from 2D keypoints. Single-camera deployments using depth estimation models (MiDaS-style monocular depth) as a proxy for 3D position produce usable but lower-confidence results.
Surface variability: Ground reaction force cannot be measured from video. The absence of force plate data means that loading inference from kinematic data requires validated biomechanical models of canine gait. These models exist in the veterinary biomechanics literature but have not been uniformly integrated into ML-based gait scoring systems. This is a known limitation that honest deployment documentation should acknowledge.
Occlusion handling: The dog's own body occludes limbs during certain phases of gait, particularly hindlimb tracking during trot when the dog is viewed from a slightly oblique angle. Temporal models handle brief occlusion better than per-frame models because they can interpolate through occluded frames using context from adjacent frames. Occlusion-robust training using synthetic data with deliberate occlusion augmentation further improves this.
Integrating Markerless Analysis into Structured Service Dog Evaluation Programs
Technology does not replace structured evaluation protocol. It augments it. The question for program directors is where in the evaluation workflow automated gait analysis adds the most value without disrupting the relational and behavioral assessment components that experienced trainers excel at.
Our recommendation, consistent with the clinical AI framework developed at TheraPetic®.AI, is a three-point integration model.
Pre-evaluation baseline capture: Before formal evaluation begins, a standardized 5-minute video walk on a consistent surface establishes a baseline gait signature for the candidate. This requires no specialist interpretation and runs automatically. The output is a candidate-specific gait profile that flags any metrics falling outside species and breed norms for veterinary follow-up before evaluation proceeds.
In-session continuous monitoring: During evaluation walks, a mounted or handler-operated camera feeds real-time inference. Gait scores are displayed to the evaluator as a continuous readout, not a post-hoc report. This changes evaluation dynamics: the trainer can observe the dog's behavioral state while the system tracks musculoskeletal state simultaneously.
Longitudinal tracking across training career: The most underutilized application is longitudinal. Comparing a candidate's gait signature at week 4, week 12, and week 24 of training reveals conditioning trajectory, injury emergence, and fatigue recovery patterns that no single evaluation session captures. The TheraPetic® Training Plus program, accessible through officialservicedog.com, incorporates structured documentation that supports this kind of longitudinal record-keeping. Integrating automated gait metrics into that record structure creates a clinically meaningful timeline for each dog.
For ADA compliance purposes, gait analysis data is not a substitute for the two-question rule that governs business interactions under Title III. It is a training and candidate selection tool. Businesses assessing whether a dog is a trained service dog under ADA.gov guidelines use behavioral criteria, not biomechanical data. The clinical value of markerless tracking sits entirely within the training program and handler-program relationship.
What the technology makes possible is a service dog candidate selection process grounded in longitudinal physiological evidence rather than point-in-time subjective impression. For programs operating at scale, that shift in evidentiary basis has meaningful implications for placement success rates, working longevity, and the welfare of the dogs being asked to serve.
The models are ready. The deployment infrastructure is maturing. The remaining work is in canine-specific training data, validated biomechanical ground truth, and program adoption. ServiceDog.AI is building in all three directions.