Spatial Sparsity-Aware Radar-LiDAR Fusion for Occupancy Grid Mapping in Automotive Driving

We address the problem of estimating a binary occupancy grid map by fusing point cloud data from radar and LiDAR sensors for automotive driving perception. To achieve this, we introduce two measurement models for fusion and formulate occupancy mapping as sparse vector reconstruction from the set of radar and LiDAR measurements. The first model, called common sparse fusion, jointly estimates a common map from all sensor measurements. The second model, called common innovative sparse fusion, assumes a shared map and an innovation component (error collector) for each sensor modality’s measurements. This approach enhances the robustness of occupancy map estimation against potential sensor mismatch and calibration errors, and inconsistencies between the two modalities. We use the pattern-coupled sparse Bayesian learning (PCSBL) algorithm to recover maps, leveraging the inherent sparsity and spatial dependencies in automotive occupancy maps. Numerical experiments on the public RADIATE dataset show that our feature-level fusion models outperform single-modality mapping and decision-level fusion models in detecting drivable areas and targets. Furthermore, statistical results with corrupted LiDAR data establish that our common innovative sparse fusion model is robust against unreliable sensor data.