CareWatch — A Computer Vision Safety Monitoring Prototype
A hackathon prototype that turns pose and scene signals into short evidence clips for human review.
The problem
Caregivers cannot continuously watch hours of footage. CareWatch explores whether computer vision can surface a small set of potentially concerning interactions for human review, with enough context to decide what happened.
This is a safety-review prototype, not a validated abuse detector. The design center is evidence and escalation, not an automated judgment.
Why pose-based heuristics
A full action-recognition model would require labeled examples that a weekend build does not have. Pose estimation offers a more inspectable first pass:
- Track people and body keypoints with YOLOv8-pose.
- Flag proximity and fast-motion interactions, or long caretaker absence.
- Send only flagged frames to a vision model for scene verification.
- Capture a short evidence clip around the event.
The tradeoff is deliberate: heuristics are explainable and cheap, but they will produce false positives and can miss subtle events.
System flow
camera frames
→ YOLOv8 pose + person tracking
→ proximity / motion heuristics
→ scene verification (hosted or local vision model)
→ 30-second rolling evidence buffer
→ push alert + SQLite event log
→ React dashboard / iOS guardian client
The repository separates responsibilities across core/detector.py, core/classifier.py, core/clip_saver.py, core/alert.py, and core/pipeline.py. Alerts have severity-based deduplication. The dashboard supports live monitoring and review of saved recordings; the iOS client is the guardian-facing surface.
Evaluation and recognition
CareWatch received Best Social Good Hack at Bitcamp 2026, verified on Devpost.
That award is external recognition of the prototype and its framing. It is not a precision/recall result. The repository does not publish sensitivity, false-positive rates, or a clinical validation study. The demo uses staged high-risk interactions and caretaker-absence scenarios.
Team and contribution
CareWatch was a team hackathon build. I worked across the pose-heuristic pipeline, evidence capture, and review workflow; teammates contributed to the surrounding application and iOS client. The repository is the source of truth for the combined implementation and does not contain a complete file-by-file authorship ledger, so this page does not attribute every module to me.
Limits
- No claim of real-world abuse-detection reliability.
- No clinical, legal, or safeguarding validation.
- False positives are expected; camera angle, occlusion, lighting, and unusual motion all matter.
- A 30-second clip is evidence for a human, not a complete account of an event.
- The project should not be deployed as a decision-making system without consent, review, and a much stronger evaluation program.