Data and code for "Uncertainty-aware LiDAR-camera robotic inspection for roadside ditch cross-section and obstruction quantification"

Measurement data and processing code accompanying the paper (Soon to be submitted) "Uncertainty-aware LiDAR-camera robotic inspection for roadside ditch cross-section and obstruction quantification" (Balaji, Aurilio, Copetti Callai, Liu, García-Hernández), Institute of Highway Engineering (ISAC), RWTH Aachen University. The dataset was collected using a Clearpath Jackal J100 ground robot equipped with a RoboSense M1 LiDAR and an Intel RealSense D455 camera to survey roadside drainage ditches. For each extracted cross-section, the pipeline reports the channel area, obstruction area, blockage ratio, a five-component uncertainty budget, and an explicit decision indicating whether the captured geometry is sufficiently complete for the measurement to be considered valid. Contents data/Reference geometry of the trapezoidal test channel and all obstruction objects; per-frame cross-section records for the manual-label, held-out, and outdoor LiDAR-only evaluations; uncertainty components; threshold-sweep results; the data underlying each figure; and manually verified segmentation masks.Licence: CC BY 4.0. code/Complete processing pipeline, including frame extraction, LiDAR-camera calibration, SegFormer segmentation, point-cloud preprocessing, cross-section and obstruction-area estimation, the five-component uncertainty budget, slope estimation, 3D reconstruction, and figure generation.Licence: MIT. Two files provide a useful starting point. code/PARAMETERS.md documents every pipeline parameter together with the source file in which it is defined. The paper uses a single fixed configuration across all recordings, obstruction objects, and test sites, allowing this claim to be independently verified. data/verify_paper_numbers.py recomputes the reported indoor evaluation statistics directly from the per-frame records and reports whether they match the published values. Raw ROS 2 recordings, full-resolution point clouds, and the trained segmentation checkpoint are not included because of their size. These materials are available from the authors upon reasonable request. The reported accuracy results do not depend on these files, as they are derived either from the manually verified segmentation masks provided in this repository or from the LiDAR-only geometric extraction pipeline.