TanodLens: Privacy-first barangay CCTV that spots emergencies, not faces

Problem Statement

Many barangays have put in CCTV, but the footage is mostly reviewed after an incident because no one can watch dozens of screens all night. Residential fires spread fast in dense informal settlements in Metro Manila, Cebu, and Davao. BFP records thousands of fire incidents a year, and the first minutes before a call is made are often lost. Road crashes on barangay roads and highways go unreported until a bystander calls. Crowded events (fiestas, Traslación, concerts, evacuation centers) risk crowd crush. At the same time, citizens are rightly worried about surveillance and face recognition under the Data Privacy Act. Barangays need real-time detection that is affordable and doesn’t violate privacy.

Proposed Solution

Tanod Lens is a small edge box (a Jetson or mini PC) that plugs into a barangay’s existing CCTV NVR. Lightweight vision models analyze each feed locally for three event types: fire and smoke (including early smoke from roofs and alleys), road incidents (collisions, a person lying on the road, a stalled vehicle), and crowd density and abnormal flow (people per square meter, sudden surges or counterflow). Before any frame is stored or sent, people and plates are blurred on the device, and no face embeddings are ever computed. When an event crosses a threshold, a short anonymized clip, location, and event type go to the tanod on-duty group (Messenger/Viber/SMS) and to a barangay dashboard. Officers can forward the alert to the BFP, PNP, or ambulance with one tap. A multimodal LLM writes a one-line incident summary and filters out false alarms such as cooking smoke or fireworks.

Target Market

• Primary users: Barangay tanods (BPATs), barangay officials, and BDRRM committees.
• Buyers (B2G): Barangays and LGUs funding CCTV upgrades, BFP fire stations, city command centers (“Safe City” programs), and event organizers.
• B2B: Subdivisions/HOAs, malls, schools, factories, and industrial parks.
• Size: 42,000+ barangays; urban barangays with existing CCTV are the beachhead market.

How AI Is Used

• Vision detection (fire/smoke, crash, fallen person): Gives 24/7 attention that human watchers can’t provide. Without AI, CCTV is only a forensic tool used after the fact.
• Crowd density estimation: Counting-based models estimate crowd density to warn before crush conditions form. Humans are bad at judging density from a screen.
• On-device anonymization model: Detects and blurs persons and plates before storage, making privacy the default.
• Multimodal LLM verification and summary: Cuts false positives (ihaw-ihaw smoke, fog, fireworks) and writes human-readable alerts in Filipino.

PLDT / Smart / ePLDT / Rokid Angle

• PLDT Enterprise managed CCTV and fiber for barangays and LGUs. Tanod Lens is an AI upsell on existing connectivity and Safe City deals.
• Smart SMS/LTE backup for alert delivery when fiber is down, and Smart IoT SIMs for edge boxes.
• ePLDT cloud for the fleet dashboard, model updates, and in-country storage of anonymized event logs.
• Rokid AR (stretch): Responding tanods see the alert location and a live anonymized feed on AR glasses while on foot patrol.

Demo Scenario (3 minutes for judges)

1. Show the dashboard with four “barangay” camera feeds (public test videos plus a live webcam pointed at the stage).
2. Play a public-domain clip of smoke rising from a roof. Within seconds, a red “SUNOG – Purok 2, Camera 3” alert pops up, and a judge’s phone gets a Viber/SMS with an anonymized 5-second clip.
3. Point out that every person in the live webcam feed is already blurred and no face data is stored. Show the privacy log.
4. Teammates simulate a crowd surge in front of the webcam. The density meter climbs to orange, then red, triggering a crowd-control advisory.
5. Show a motorcycle crash test clip and a one-tap “Forward to BFP/911” action. Close: “Eyes that never sleep, and never stare at your face.”

30-Hour Build Plan

Suggested stack: Python, OpenCV, YOLOv8 (fire/smoke, vehicles, persons), CSRNet or a similar crowd-counting model, ONNX Runtime/TensorRT on Jetson or laptop, a multimodal LLM (Gemini/GPT-class), FastAPI, React dashboard, SMS/Viber/Messenger APIs. Public data / APIs: Hugging Face and Roboflow Universe fire/smoke datasets, D-Fire dataset, ShanghaiTech/UCF crowd-counting datasets, public dashcam/CCTV crash datasets (e.g., CADP), BFP fire statistics, OpenStreetMap for camera geolocation.

Risks & Privacy Notes

• False alarms (cooking smoke, fog): Require detection over multiple frames, use the LLM verification step, and tune thresholds for each camera. A human tanod confirms before escalating to BFP.
• Missed detections at night or in rain: Use IR-capable cameras where available, and position Tanod Lens as an aid, not a replacement for 911.
• Data Privacy Act 2012: Purpose limitation (safety events only), no biometric processing, on-device blurring, short retention (e.g., 7 days for event clips), visible CCTV signage, and a Privacy Impact Assessment with the barangay as personal information controller.
• Mission creep: No face recognition or person tracking by design or in the license terms, so the product can’t be turned into a surveillance tool.
• Clause 24 compliance: Use only public datasets and staged team footage. Only anonymized frames ever reach cloud AI APIs.

AI Use Disclosure (draft)

Tanod Lens uses YOLO-class detectors for fire, smoke, vehicles, and persons, a crowd-counting CNN, an on-device anonymization model, and a multimodal LLM to verify and summarize alerts in Filipino. AI coding assistants helped with code scaffolding and drafting documentation. Every alert is reviewed by a human tanod before escalation, and the team reviewed all AI-assisted code and text.

Rubric Scoring