Drowning Detection in Pools using CNN

Drowning is silent and fast: a struggling swimmer cannot call for help, and a lone lifeguard cannot watch every corner of a busy pool at once. Automated pool monitoring is an active research area, but the honest framing for a student project is an alert aid, not a safety device — a vision system that flags suspicious motion patterns for a human to check, with documented failure modes and no life-safety claims. This project builds that aid: OpenCV frame differencing produces motion maps from pool camera feeds, pose cues from MediaPipe track swimmer movement, and a lightweight classifier scores motion sequences over a time window for distress-like patterns such as prolonged struggling or bobbing without net displacement. The alert logic requires sustained patterns so passing splashes and normal play do not become alerts, and every alert candidate is logged with a timestamp and snapshot for human review. The evaluation procedure runs on buyer-staged footage with volunteers acting out scenarios, because real drowning footage is neither available nor ethical to collect.

1. Pool camera frames are sampled and background subtraction with frame differencing produces per-frame motion maps.
2. Motion regions are filtered by area and persistence; static water ripple noise is suppressed with temporal smoothing.
3. MediaPipe pose landmarks track swimmers, measuring vertical bobbing and displacement over the time window.
4. A lightweight classifier scores each motion sequence for distress-like patterns — sustained struggling motion without net displacement.
5. Patterns persisting across the time window become confirmed alert candidates logged with timestamp and snapshot; isolated blips are discarded.
6. The notebook runs the buyer-staged evaluation: labeled scenarios are replayed through the pipeline and the alert rule's precision and recall are computed.

Tech stack:

• Python 3.10, OpenCV (frame differencing, motion maps)
• PyTorch (motion-sequence classifier training)
• MediaPipe pose (body landmarks and motion cues)
• NumPy, pandas (motion features and alert logs)
• Jupyter notebook (staged-footage evaluation procedure)
• Matplotlib (motion timelines and evaluation plots)
• Flask demo app with camera-feed upload interface

• Frame-differencing motion maps from pool camera feeds via OpenCV
• MediaPipe pose cues tracking swimmer movement across frames
• Motion-sequence classifier scoring struggle-like patterns over a time window
• Distress heuristics: prolonged splashing, bobbing without displacement
• Time-window alert rule suppressing single-blip false alerts
• Timestamped alert log with annotated frame snapshots
• Buyer-run evaluation procedure on staged footage (volunteers acting out scenarios)
• Notebook computing the alert rule's precision and recall on the staged set

• Complete source code (motion pipeline, pose cues, classifier, alert logic, demo app)
• Jupyter notebook with the buyer-run staged-footage evaluation procedure
• Project report PDF (background, motion methodology, evaluation, honest safety limits)
• PPT presentation for final review
• Viva Q&A preparation document (frame differencing, pose cues, time-window alerts, evaluation design)
• Setup guide (environment, camera placement, staging the evaluation footage, running the pipeline)

• This is a prototype only, NOT a safety device: it never substitutes lifeguards or supervision, and it must not be deployed where people depend on it — not for real deployment.
• False negatives are the critical failure mode — a still, submerged victim produces little motion, so the system can miss exactly the cases that matter most; the report states this limitation explicitly.
• Busy pools, glare, reflections and swimmers outside the camera's view degrade detection; evaluation is on buyer-staged footage with volunteers, never on real incidents, for ethical reasons.