Abstract
An edge road hazard classifier processes monocular camera video to detect and localize physical road surface anomalies. The system calculates the relative distance offset of a detected hazard using inverse perspective mapping (IPM). To estimate hazard depth, the system combines structure from motion (SfM) and a monocular depth estimation network. The hazard classifier bundles these measurement values into a metadata packet for transmission to a dynamic map ingestion engine. The map ingestion engine normalizes differences in sensor noise with a spatial covariance error matrix and groups overlapping reports using spatiotemporal HDBSCAN (ST-HDBSCAN). An extended Kalman filter (EKF) synthesizes the grouped reports to determine a final geographic coordinate. A temporal decay engine updates hazard statuses by detecting swerve maneuvers and analyzing visual frustum coverage. Keywords include IPM, SfM, and ST-HDBSCAN.
Keywords: inverse perspective mapping (IPM), IPM geometric calibration, structure from motion (SfM), monocular depth estimation network, spatial covariance error matrix, spatiotemporal HDBSCAN (ST-HDBSCAN), extended Kalman filter (EKF), visual frustum, swerve maneuver, temporal decay engine, dynamic map ingestion engine
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Gupta, Richa, "Monocular Edge-Based Road Hazard Detection with Covariance-Weighted Spatiotemporal Clustering and Frustum Temporal Decay", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12245