Autonomous vehicle self-localization based on probabilistic planar surface map and multi-channel LiDAR in urban area

Autonomous vehicle self-positioning based on 3D light detection and ranging (Lidar) has become popular recently due to disadvantages of global navigation satellite system (GNSS) in urban areas. As LiDAR-based simultaneous localization and mapping (SLAM) methods suffer from error accumulation, state-of-the-art approaches match the point cloud data acquired by LiDAR to the priori known 3D point cloud map to obtain the position of the vehicle within the map. However, 3D point cloud map is very expensive to store and download as it contains an enormous amount of data even for a small area (around 300 million points per km
²
). In this study, rather than using 3D point cloud directly as a map, we focused on the planar surfaces which are mostly available in urban areas, easy to extract, and at the same time clearly observable by LiDAR. Therefore, in our proposed map, we extract the planar surfaces from the 3D point cloud and calculate its uncertainty (deviation) and store them as a prior map. Accordingly, in this map, we can abstract several thousands of points by only one plane. As a result, we can extremely shrink the map size (25 million points to around 1000 planes). Later in the localization phase, we reconstruct Gaussian mixture model for each planar surface based on previously stored deviation, and match LiDAR data to it to obtain the precise location of the vehicle. Experiments conducted in one of the urban areas of Tokyo show that even though we extremely shrank the map size, we could preserve the mean error of the localization less than 43cm comparing to other point cloud based methods.