The Problem and Its Impact Traditional flood monitoring often fails because it is either too passive or relies on expensive, dedicated hardware. For many citizens, the issue isn't just knowing that "it is raining," but understanding exactly when the water level reaches a critical threshold that necessitates immediate action. Current systems frequently require people to proactively check apps or news reports, which can lead to delayed responses during rapid-rise events. Our project addresses this by shifting the burden from the citizen to the infrastructure; instead of citizens checking the system, the system proactively reaches out to them when a verified threat is detected in their specific area.
The Solution: An Engineering Approach to Visual Data We transform passive visual streams into machine-readable environmental data using a rigorous 5-step pipeline: SEE, MEASURE, VERIFY, DECIDE, and WARN.
- SEE: The system utilizes existing CCTV infrastructure as its primary source of observation.
- MEASURE: Moving beyond simple "flood" vs. "no flood" classification, the system uses visual geometry and calibrated scenes to estimate the exact water level in centimeters.
- VERIFY: This is the "engineering soul" of the project. We don't trust a single frame because visual noise, reflections, or obstructions can cause errors. We analyze temporal consistency (history over time) and visual quality to produce a Confidence Score.
- DECIDE: A risk engine combines the current water level, the trajectory (rate of rise), and the confidence score to determine if the situation is truly dangerous.
- WARN: When high-confidence danger is identified, a localized, geofenced alert is triggered for the specific affected area.
What Makes VFI Stand Out
- Quantitative Instrumentation: We treat the camera as a sensor providing real-time measurements in centimeters, not just an image classifier.
- Confidence-Aware Logic: Our system recognizes that a warning system is most dangerous when it is "confidently wrong". By quantifying Reliability (Confidence) separately from Danger (Risk), we ensure that alerts are only issued when the data is both critical and trustworthy.
- Safe Behavior Philosophy: If visibility is poor or the waterline is obscured, the system is designed to report "measurement unavailable" rather than inventing numbers, preventing false panic.
- Proactive & Localized: Unlike global dashboards, our system issues area-specific warnings through a "one-to-many" trigger mechanism, targeting only the people in the immediate path of the water.
Challenges and Key Learnings The greatest challenge was overcoming the unreliability of individual visual frames. We learned that temporal context is essential; a water level that is 60cm and stable is fundamentally different from a level that is 60cm and rapidly rising. We also discovered that for a system to be used in public safety, it must be conservative—it must account for its own uncertainty and only escalate warnings when the evidence across time is consistent and clear.
Log in or sign up for Devpost to join the conversation.