LiDAR-Supervised First-Return Occupancy Grid Mapping for Automotive Radar on the RaDelft Dataset
Y. Pang (TU Delft - Electrical Engineering, Mathematics and Computer Science)
F. Fioranelli – Mentor (Microwave Sensing, Signals & Systems)
S. Yuan – Mentor (Microwave Sensing, Signals & Systems)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Automotive radar is robust to adverse weather and illumination conditions, but its sparse, noisy, and not always spatially accurate measurements make dense occupancy estimation challenging. This thesis investigates light detection and ranging (LiDAR)-supervised radar-based first-return occupancy grid mapping on a polar range–azimuth grid. Radar is used as the sole input modality during inference, while synchronized LiDAR point clouds are processed into radar-aligned supervision representing visible free space, the nearest occupied boundary, and the unknown region behind it.
The proposed framework formulates occupancy estimation as an azimuth-wise first-return prediction problem rather than independent cell-wise classification. The complete range–Doppler–azimuth (RDA) radar cube is processed using a Range–Azimuth-Preserving Doppler Spectrum Encoder with parallel multi-scale Doppler-only branches, followed by a U-Net 3+ prediction backbone. The predicted first-return distribution is rendered into a visibility-consistent three-state occupancy grid. Training combines a Lovasz-based occupancy objective with a smoothed cumulative distribution function (CDF)-L1 boundary loss. Radar-aware azimuth weighting and causal temporal feature fusion are further investigated as controlled extensions.
Experiments are conducted on the RaDelft dataset using a scene-level train, validation, and test split. The proposed single-frame baseline achieves a free-space intersection over union (IoU) of 0.79, an unfree-space IoU of 0.82, a mean IoU (mIoU) of 0.80, and a first-return mean absolute error (MAE) of 5.38 m. It outperforms an odometry-compensated multi-frame inverse sensor model and a RaDelft-adapted PolarNet reference taken from the literature. Additional experiments show that retaining the complete Doppler spectrum is more effective than fixed mean or maximum aggregation. The proposed Doppler encoder provides accuracy comparable to a reference encoder from the literature, but at lower computational cost.
Overall, the results demonstrate that structured first-return prediction from full Doppler-resolved radar measurements provides an effective approach to LiDAR-supervised polar occupancy mapping, with the single-frame full-RDA model offering the most favourable accuracy–efficiency balance among the evaluated full-RDA configurations.