Road Boundary Estimation Using Sparse Automotive Radar Inputs

Low-cost millimeter wavelength automotive radar can work effectively under low visibility or low reflection con- ditions caused by lighting, weather, pollution, or object surface properties when a camera or a lidar may fail. It can serve as a fallback solution to improve safety in autonomous driving. However, after filtering, radar signals tend to be sparse and noisy which poses new challenges in scene understanding. This paper presents a new approach to detecting road boundaries based on sparse radar signals. We model the roadway using a homogeneous model and derive its conditional predictive model under known radar motion. Using this predictive model and modeling radar points using a Dirichlet Process Mixture Model, we employ Mean Field Variational Inference (MFVI) to derive an unconditional road boundary model distribution. To generate initial candidate solutions for the MFVI, we develop a custom Random Sample and Consensus (RANSAC) variant to propose unseen model instances as candidate road boundaries. For each radar point cloud we alternate the MFVI and RANSAC proposal steps until convergence to generate the best estimate of all candidate models. We select the candidate model with the minimum lateral distance to the radar on each side as the estimates of the left and right boundaries. We have implemented the proposed algorithm and it has shown satisfactory results. More specifically, the mean lane boundary estimation error is not more than 11.0 cm.