Sensor-Free Millimeter-Level Water Depth Estimation for Flood-Aware Route Optimization

Abstract

Urban areas frequently face the compounded challenges of severe traffic congestion and localized flash flooding, rendering traditional navigation systems inadequate. Existing vision-based approaches either address traffic monitoring or flood detection in isolation, and none achieves the millimeter-level physical precision required for safe vehicle routing. This study proposes a novel vision-based sensing framework for dynamic route optimization that simultaneously estimates traffic density and physical water depth utilizing solely standard traffic camera feeds, requiring no additional sensors or infrastructure. The framework integrates CSRNet for vehicle density estimation and BiSeNet for real-time flood segmentation, achieving an mIoU of 95.1%. This is coupled with a baseline-referenced contour depth method that leverages static anchor objects, extracted by SAM 3, to neutralize scale ambiguity. Furthermore, a two-step statistical filtering pipeline combining Median-IQR outlier rejection and spatial coherence enforcement stabilizes depth measurements under adverse conditions. This approach achieves millimeter-level accuracy with a Mean Absolute Error (MAE) of 1.53 mm, substantially outperforming state-of-the-art monocular depth estimators and Vision-Language Models. Finally, these continuous metrics are incorporated into a multi-objective cost function, optimized via grid search, and solved using the A* pathfinding algorithm to dynamically recalculate safe trajectories. Experimental results demonstrate the system's capability to redirect vehicles away from hazardous intersections while minimizing travel delays, providing a robust, sensor-free solution for Intelligent Transportation Systems (ITS) to enhance urban mobility and safety during extreme weather events.