Offline Covariance Prediction for Lidar-Based Map-Matching in Autonomous Systems

This paper introduces a concept for map-based characterization of lidar positioning errors. A map-based model can account for spatial variations in lidar positioning performance, including variations in the types of terrain the lidar visualizes and matches to a high-definition map (HD map). Our approach for map-based error characterization relies on a two-step positioning algorithm. The two-step algorithm includes a coarse positioning step using scan context descriptors, as well as a refined step using normal-distribution transform (NDT) scan matching. This two-step approach is practical in that it reduces sensitivity to the initial guess needed for refined scan matching. Initial-condition sensitivity can otherwise cause problems with scan-matching convergence and introduce biases in statistical error characterization. After describing our map-based error characterization approach, we apply the method to a representative highway dataset. Our analysis shows that the map-based error characterization approach provides a reasonable characterization of errors in a test data set. We also provide evidence that lidar-positioning error distributions exhibit heavier-than Gaussian tails.