Monocular Curb Edge Detection via Robust Geometric Correspondences

Advanced driver-assistance and autonomous systems require perception that is both robust and affordable. Monocular cameras are promising due to their ubiquity and low cost, yet detecting abrupt road surface irregularities such as curbs and bumps remains challenging. These sudden road gradient changes are often only a few centimeters high, making them difficult to detect and resolve from a single moving camera. We hypothesize that stable image-based homography, derived from robust geometric correspondences, is a viable method for predicting sudden road surface gradient changes. To this end, we propose a monocular, geometry-driven pipeline that combines transformer-based feature matching, homography decomposition, temporal filtering, and late-stage IMU fusion. In addition, we introduce a dedicated dataset with synchronized camera and ground-truth measurements for reproducible evaluation under diverse urban conditions. We conduct a targeted feasibility study on six scenarios specifically recorded for small, safety-relevant discontinuities (four curb approaches, two speed bumps). Homography-based cues provide reliable early signatures for curbs (3/4 curb sequences detected at a 5 cm threshold). These results establish feasibility for monocular, geometric curb detection and motivate larger-scale validation. The code and the collected data will be made publicly available.