Map-based registration of multi-LiDAR infrastructure sensor setups for real-time applications

Abstract

The use of infrastructure sensors has proven to be an effective method for recording large traffic datasets. In order to achieve a high data quality for such datasets, an accurate extrinsic calibration of the sensors is essential. While previous approaches are not fast enough or do not fit the use-case of infrastructure sensing, we present a method that is capable of registering LiDAR sensors to a digital map in real-time. We use a mobile infrastructure sensor setup consisting of two measurement units, which we position at different locations with real road traffic. Each measurement unit is equipped with two LiDARs at a mounting height of eight meters. Experiments have shown that our method can perform continuous registration of up to four 128-layer LiDARs. With a voxel size of 0.05 meters used, our method achieves a root-mean-square error (RMSE) of 0.13 meters at an average processing time of 0.05 seconds per frame.