Fully Automatic Multi-LiDAR Calibration For Self-Driving Cars

Abstract

This paper addresses automatic vehicle-relative calibration of LiDAR sensors, which is essential for self-driving cars, e. g., for accurate localization, SLAM-based mapping, and sensor data fusion. While various semi-automatic techniques exist for the calibration of environment sensors, they still require a considerable amount of manual work or an expensive setup. To that, this paper presents a new algorithm for finding the mounting poses of multiple low-density LiDAR-sensors on the vehicle while minimizing manual effort. In particular, the algorithm provides co-calibration as well as extrinsic calibration relative to the IMU installed on the vehicle. The calibration process only requires some recorded data from the sensors but no special environment to compute the accurate poses. In contrast to other algorithms that register point clouds to each other, this algorithm uses a newly developed map-to-map matching method increasing the robustness of optimization processes with regard to noisy data. It works by maximizing all sensors’ consensus regarding a map simultaneously created from the recorded data. We evaluated our algorithm using a car with six mounted LiDAR-sensors and recordings from different scenarios in real urban traffic and within a testing area. Our algorithm calibrates the sensors with an expected error of less than 5 cm and less than 0.5 degrees. A key contribution is its robustness against highly noisy initial sensor configurations and its accuracy regarding the angular calibration. A visualization of the recorded data also shows calibration improvements of the mounting poses over the manual measurements.