Thesis · 2024 · ML / CV lead
SaveMe
Computer-vision drowning detection for backyard pools — a YOLO pipeline trained on a hand-annotated dataset that turns an ordinary camera into an early-warning signal.
PythonPyTorchOpenCVYOLO
The problem
Residential pools rarely have lifeguards. The goal was an early-warning signal from an ordinary camera feed — not a wearable.
Approach
OpenCV ingests the camera feed frame by frame; a YOLO detector locates swimmers; the pipeline watches distress-associated posture and stillness over time.
Because no suitable dataset existed, training ran on a custom-annotated set of residential-pool footage.
Decisions
Annotating a custom dataset instead of reusing one
Public swimming datasets are filmed at competition pools and angles. A model trained on them generalises badly to a backyard camera, so the dataset had to match the deployment conditions.