Az útfelület normálvektorának predikciója és gyakorlati alkalmazásai = Prediction of road surface normal vector and its practical applications

Markó, Norbert and Ballagi, Áron and Szirányi, Tamás (2026) Az útfelület normálvektorának predikciója és gyakorlati alkalmazásai = Prediction of road surface normal vector and its practical applications. SCIENTIA ET SECURITAS, 6 (4). pp. 354-360. ISSN 3057-9759

Preview

Text 112-article-p354.pdf - Published Version Available under License Creative Commons Attribution. Download (968kB) | Preview

Abstract

Tanulmányunk összefoglalja korábbi cikkeink eredményeit az útfelület normáljának monokuláris kamerából történő becsléséről. A vizsgált probléma a lejtők, emelkedők és hirtelen átmenetek okozta torzulások kezelése. Módszerünk képpáralapú homográfián alapul, amelyet a késői IMU-fúzióval világkoordinátába forgatunk, illetve időben SLERPszűréssel stabilizálunk. Az algoritmust a PandaSet, KITTI és saját GradeSet adatkészleteken értékeltük ki, ahol újrakalibrálás nélkül, következetesen alacsony normál- és pitch-hiba-értékeket kaptunk a referenciához és az előző state-of-the-art algoritmushoz képest. A kutatásaink során kidolgozott megközelítés robusztus, jól általánosítható alapot ad útfelület normálisának kameraalapú becsléséhez, ami számos további gyakorlati alkalmazásban jól hasznosítható. | This paper summarizes our previous results on estimating the road surface normal from a monocular camera. We address distortions caused by slopes, grades, and abrupt transitions where projection errors grow quickly. A one degree tilt error at 50 m implies almost one meter vertical discrepancy, which can mislead free-space inference, 3D deprojection and obstacle reasoning. By explicitly recovering the road plane orientation we provide a strong geometric prior for camera-first perception. The pipeline consumes two consecutive frames, enhances road texture with contrast-limited adaptive histogram equalization, and computes dense correspondences with a lightweight transformer matcher (EfficientLoFTR). Robust homography is then estimated with MAGSAC over a road-only region of interest set near a reaction distance of roughly six meters. The homography is normalized with the intrinsic calibration to isolate geometry and improve numerical stability, then decomposed into rotation, translation and plane normal. Because matches are constrained to the road, the recovered normal describes the ground plane in the camera frame. We rotate this normal into a physically meaningful world frame through late fusion with IMU-based odometry, and stabilize the time series with spherical linear interpolation across unit quaternions. SLERP delivers smooth, unit-length updates without the parameter sensitivity and matrix inversions of Kalman-style filters, while remaining efficient for embedded deployment. We evaluate on PandaSet, KITTI, and our own GradeSet, a dataset that targets dynamic grade changes and steep ramps. A single set of parameters is used across all data, yet the method maintains consistently low errors. Averaged over challenging segments we measure a normal error of 1.18 degrees and a pitch error of 0.75 degrees, compared with 3.15 and 1.72 degrees for a strong reference. The approach handles low-texture asphalt, lane-marking scarcity and rapid grade transitions, and it does not require dataset-specific recalibration. Beyond accuracy, the explicit surface orientation enables precise inverse perspective mapping, strengthening freespace detection from a single moving camera. Practical payoffs include adaptive speed control on heavy ground vehicles, which benefits from a stable estimate of slope and pitch for better comfort, energy use and safety. The predicted normal also supports continuous compensation of camera extrinsics relative to the road by combining an initial vanishing-point calibration with our per-frame normal, reducing long-term drift and improving consistency in camera-only stacks. The result is a robust, generalizable and resource-efficient basis for road-surface normal estimation. We also outline a forward path toward piecewise-plane modeling and extension to unstructured off-road terrain, where richer features and surface models can carry the same late-fusion and temporal-smoothing principles to more varied geometry while preserving the efficiency advantages of cameras.

Actions (login required)

Edit Item