Semantic LiDAR Point Cloud Mapping and Cloud-Based SLAM for Autonomous Driving

Abstract

This paper presents a strategy towards reliable LiDAR-based navigation for applications as self-driving cars in urban canyons with limited GNSS availability. The general idea is a cloud service which provides updated, large-scalable, and geo-referenced point cloud maps. Vehicles can download snapshots on demand and use them for map-based positioning. On the mapping side of the network, multiple connected cars collect data and share them with the cloud service which performs all the data fusion and mapping tasks. While mapping and localization are state-of-the-art, this SLAM-as-a-Service idea now allows to scale SLAM-based navigation into arbitrary large areas, and it solves issues in positioning error accumulation and in the very first drives in a previously unknown environment. The implementation of this approach is tested with an experimental vehicle driven in real urban traffic, and it can be shown that state estimation is improved in relation to GNSS. With good mapping, positioning returns 15 cm geodetic accuracy on a smooth trajectory without GNSS-typical position jumps.