Service dog public access assessment is one of the most contested verification challenges in disability accommodation law. Under the Fair Housing Act and the Air Carrier Access Act, housing providers and transportation operators are permitted to ask two specific questions about a service animal, but they cannot legally require formal certification or demonstration of trained tasks. That regulatory gap creates a real problem for AI engineers and clinical teams trying to build scalable, legally defensible verification infrastructure.
At TheraPetic® Healthcare Provider Group, our clinical and engineering teams have spent considerable time evaluating whether canine pose estimation models can close part of that gap. The short answer: they can surface meaningful behavioral signal. The longer answer involves understanding exactly what current architectures like DeepLabCut and SLEAP can and cannot do when applied to real-world public access video. This article reviews the technical architecture of both frameworks, identifies the behavioral markers most reliably captured through video analysis and addresses the hard limits of single-camera assessment that any honest deployment must account for.
Why Video-Based Assessment Matters for Service Dog Verification
The public access test, as described by organizations like Assistance Dogs International, evaluates a dog's ability to remain calm, focused and non-reactive across a range of real-world environments. It is a behavioral assessment, not a paper credential. That behavioral nature makes it a natural candidate for computer vision analysis.
Traditional verification models rely on self-reported handler attestation, often supplemented by documentation from a Licensed Clinical Doctor confirming the handler's disability-related need. That clinical layer is necessary and legally sound. What it cannot do is independently evaluate whether the animal accompanying a patient meets behavioral standards for public access. A Licensed Clinical Doctor can assess human need. They are not trained as canine behavior specialists.
Video-based behavioral analysis fills that specific gap. If a handler submits a short standardized video of their dog performing public access scenarios, a pose estimation pipeline can extract joint trajectories, flag stress postures and measure behavioral consistency across trials. That output can then be reviewed by a canine behavior specialist or used as structured input for clinical documentation. At TheraPetic®, our HANK AI infrastructure at verify.mypsd.org includes intake workflows designed to receive and triage exactly this kind of behavioral evidence.
The goal is not to automate the decision. The goal is to produce structured, reproducible behavioral data that reduces reliance on purely subjective attestation.
Canine Pose Estimation Architectures: DeepLabCut and SLEAP
Two frameworks dominate research-grade canine pose estimation as of 2026: DeepLabCut and SLEAP. Both were originally developed for neuroscience applications, where precise tracking of animal movement is essential for behavioral phenotyping. Both are now being evaluated in applied veterinary and service animal research contexts.
DeepLabCut
DeepLabCut, developed by the Mathis Group and published in Nature Neuroscience (Mathis et al., 2018, doi:10.1038/s41593-018-0209-y), uses a ResNet backbone with deformable convolutional layers to predict keypoint locations on individual animals from labeled training frames. The landmark contribution of DeepLabCut was demonstrating that a small number of labeled frames, typically 50 to 200, could achieve markerless pose tracking competitive with traditional marker-based systems.
For canine applications, a standard skeleton might label 17 to 22 keypoints: snout, ears, base of neck, shoulder joints, elbow joints, wrist joints, hip joints, stifle joints, hock joints, tail base and tail tip. With sufficient training data across coat types and lighting conditions, DeepLabCut achieves sub-pixel accuracy on constrained laboratory video. Real-world outdoor or retail environment video introduces significantly higher error rates due to occlusion, variable lighting and camera motion.
SLEAP
SLEAP, developed at the Janelia Research Campus and published in Nature Methods (Pereira et al., 2022, doi:10.1038/s41592-022-01426-1), extends pose estimation to multi-animal scenarios. This matters enormously for public access assessment, where the dog and handler must be tracked simultaneously in crowded environments. SLEAP uses a top-down or bottom-up approach depending on the number of animals, and its part affinity field architecture allows it to associate detected keypoints with specific individuals even under occlusion and near-overlap conditions.
For service dog verification video, SLEAP's multi-instance tracking enables simultaneous body language analysis of the dog and the handler. Handler posture, leash tension geometry and spatial proximity patterns can all be derived from the dual-skeleton output.
Behavioral Markers Detectable Through Pose Estimation
The practical value of any pose estimation pipeline depends on whether the extracted keypoint data maps onto behaviorally meaningful signals. In the context of public access assessment, several marker categories have documented validity in canine behavior science.
Gaze and Head Orientation
Head angle relative to the handler, estimated from the snout-to-ear-base vector, provides a proxy for attentional focus. A service dog in a well-trained working state maintains frequent handler-check behavior, characterized by periodic reorientation of the snout toward the handler during heel position. Persistent forward-facing head orientation combined with ear flattening or forward ear carriage can indicate environmental reactivity. DeepLabCut ear keypoint tracking has been used in published research to estimate arousal state via ear position angle.
Tail Carriage and Movement
Tail position relative to the lumbar baseline, captured via tail base and tail tip keypoints, is a reliable low-inference behavioral signal. Tail carriage above the topline in certain breeds indicates heightened arousal. Tail tucking, defined as tail tip trajectory falling below the hock joint, is a validated stress indicator. Tail flagging, rapid lateral oscillation at high amplitude, is associated with excessive excitement inappropriate for public access contexts. All three patterns are geometrically extractable from frame-by-frame keypoint data.
Body Tension and Spinal Alignment
Spinal curvature, estimated from the neck-shoulder-hip-tail base chain, differs measurably between relaxed working posture and stress-loaded posture. A dog carrying tension through the shoulders and lumbar region will show a compressed topline with elevated scapular displacement. This is detectable as deviation from a neutral spinal angle baseline established from the first seconds of video when the dog is stationary.
Leash Geometry as a Behavioral Proxy
When handler keypoints are tracked simultaneously via SLEAP, the geometric relationship between the handler's hand position and the dog's collar can serve as a proxy for leash tension. A taut leash arc indicates the dog is pulling, lunging or hanging back, all of which are disqualifying behaviors in formal public access standards. A loose leash arc with consistent spatial distance maintenance indicates appropriate heel behavior. This derived feature requires no additional sensor hardware beyond the video feed.
Environmental Reactivity Events
Sudden velocity spikes in keypoint trajectories, specifically the snout and shoulder keypoints, signal startle or lunge events. Frame differencing on keypoint coordinates allows automated flagging of high-acceleration movement episodes. Combined with video timestamp metadata, these events can be cross-referenced with known environmental stimuli in the video scene, passing pedestrians, shopping carts, other animals, to distinguish appropriate alerting behavior from inappropriate reactivity.
The Limits of Single-Camera Assessment
Single-camera video assessment is the realistic submission format for any scalable handler-submitted verification workflow. Expecting handlers to set up calibrated multi-camera rigs is not operationally feasible. This constraint introduces hard technical limits that any honest deployment of computer vision for service dog assessment must acknowledge.
Depth ambiguity is the primary problem. A single 2D camera cannot resolve the difference between a dog that is actually pulling forward and a dog that is simply positioned at an angle that makes the leash appear taut. Pose estimation operates in image space, not world space. Without stereo camera pairs or depth sensors, all keypoint coordinates are pixel coordinates, not metric coordinates. This means absolute distance measurements, such as the precise spatial gap between handler and dog, cannot be computed. Only angular relationships and relative proportions are available.
Occlusion compounds this. In any real-world public access scenario, the dog will move behind the handler, behind retail fixtures, behind other pedestrians. When keypoints are occluded for more than 10 to 15 consecutive frames, modern pose estimation models either hallucinate plausible keypoint positions through temporal smoothing or produce confidence-weighted dropouts. Neither is ideal for behavioral assessment. DeepLabCut's likelihood filter allows flagging of low-confidence frames, but in a crowded environment, a meaningful fraction of frames may fall below the threshold.
Breed-specific model drift is another documented limitation. Most open-source canine pose estimation models were trained predominantly on short-haired, medium-sized dogs. Breeds with dense coats, very large frames or extreme morphology, such as Newfoundlands, Dachshunds or Chow Chows, show significantly higher keypoint localization error due to distribution shift from training data. A service dog population that skews toward Labrador Retrievers and Golden Retrievers will produce more reliable results than one including a diverse morphological range. This bias must be disclosed and actively mitigated through breed-stratified training data collection.
Lighting and camera quality variation is a practical reality in handler-submitted video. Consumer smartphone video in a dimly lit apartment hallway is qualitatively different from well-lit outdoor retail footage. Current pose estimation models degrade meaningfully in low-light or high-motion-blur conditions. Any production deployment needs a video quality gate that rejects or flags submissions below minimum technical thresholds before running inference.
Integration with Clinical Verification Workflows
At TheraPetic® Healthcare Provider Group, the position of computer vision output in the verification workflow is explicitly adjunctive. Our Licensed Clinical Doctors assess the human patient's disability-related need. HANK AI coordinates intake, document triage and anomaly flagging. Canine behavioral analysis, when included, produces structured behavioral reports that are reviewed alongside clinical documentation, not substituted for it.
In practical infrastructure terms, this means the pose estimation pipeline runs asynchronously after video submission. Output is structured as a behavioral summary JSON object that includes flagged events, aggregate behavioral scores by category and confidence-weighted reliability estimates. That JSON feeds into the clinical review interface at verify.mypsd.org, where a reviewer can inspect specific timestamped clips alongside the pose overlay visualization.
This architecture keeps the human reviewer in the decision loop while dramatically reducing the cognitive load of manual video review. Watching 90 seconds of handler-submitted video and forming a reliable behavioral judgment is harder than it sounds. Reviewers tend to anchor on the first and last 10 seconds. Pose estimation flags events across the entire clip, including mid-clip reactivity episodes that a tired human reviewer might miss.
FHIR R4 structured data standards, specifically the Observation resource, are used to encode behavioral findings in a format compatible with the broader clinical record. This is important for audit trail integrity and HIPAA-compliant data governance, managed through mydatakey.org for patient-controlled record access.
Ethical and Regulatory Considerations in AI-Assisted Service Animal Screening
The deployment of any AI-assisted tool in service animal verification carries significant civil rights implications. The Fair Housing Act and the Americans with Disabilities Act place strict limits on the burden that can be imposed on individuals with disabilities seeking accommodation. Any computer vision assessment that functions as a gatekeeping mechanism, rather than a documentation support tool, risks running afoul of HUD guidance on reasonable accommodation requests.
This is not a hypothetical concern. A system that systematically rejects video submissions from handlers with certain dog breeds, or that flags behavioral patterns more common in dogs trained in non-Western traditions, introduces algorithmic bias with direct civil rights consequences. Equalized odds analysis across breed, handler demographics and training methodology is a minimum fairness evaluation requirement before any such system approaches production deployment.
The FDA's Software as a Medical Device framework is relevant to the extent that behavioral output informs clinical documentation. Our legal and clinical teams at TheraPetic® Solutions Inc. have structured the HANK AI pipeline to position behavioral video analysis as clinical decision support, not as an autonomous determination, specifically to remain within the lower-risk device classification that does not require premarket approval.
Transparency with handlers is non-negotiable. Any submission workflow that uses AI analysis of submitted video must disclose that analysis, explain what is being measured, describe how output is used and provide a human review pathway for disputed assessments.
Future Directions: Multi-Modal Canine Behavior Analysis
The single-camera pose estimation approach reviewed here represents the current practical ceiling for handler-submitted video assessment. Meaningful accuracy improvements will require moving toward multi-modal data fusion.
Acoustic analysis of environmental soundscapes paired with video, allowing correlation of specific sound stimuli with behavioral responses, is a near-term extension worth engineering attention. Models that can tag the timestamp of a sudden loud noise and correlate it with a detected startle response keypoint trajectory would produce far more interpretable behavioral output than pose data alone.
Wearable inertial measurement units on the dog's harness or collar would provide the depth and metric-space movement data that single-camera systems cannot. Integration of accelerometer and gyroscope data with video keypoint streams is computationally tractable with current edge inference hardware. The barrier is handler compliance with wearable submission requirements, not technical feasibility.
Foundation models trained on large-scale animal behavior video datasets are an emerging area that deserves close monitoring. Work by groups at the Broad Institute and DeepMind on behavioral foundation models suggests that transfer learning from large diverse training corpora could substantially reduce the breed-specific distribution shift problem that currently limits canine pose estimation reliability.
At servicedog.ai, the TheraPetic® network's companion AI platform, longitudinal behavioral data collection across verified service dog handler pairs is an active infrastructure priority. Longitudinal data, tracking the same dog across multiple video submissions over months and years, would allow drift detection, training consistency evaluation and outcome research that cross-sectional single-submission analysis cannot support.
The intersection of computer vision, canine behavioral science and disability accommodation law is genuinely new territory. The frameworks exist. The behavioral science exists. Building the bridge between them in a way that is technically rigorous, clinically valid, legally defensible and equitable for people with disabilities is the work in front of us.