Dynamic lidar extrinsic calibration with automatic scene selection

LiDAR is one of the indispensable sensors for high-level autonomous driving, providing accurate 3D environment data for autonomous vehicles. Typically, autonomous vehicles are equipped with multiple LiDAR sensors to expand the field of view. Calibration is a critical step in using these sensors. It was shown that accurate calibration of LiDAR can be achieved using a SLAM (simultaneous localization and mapping) algorithm.

At the front end of this algorithm, point cloud registration is a critical step. Currently, the mainstream point cloud registration algorithms include ICP and NDT, etc. These algorithms show different performances in different scenarios, and the selection of an appropriate point cloud registration method can improve the accuracy and speed of LiDAR sensor calibration. Meanwhile, the existing algorithms requires a specific driving route in some dedicated test environments (e.g. figure-8 drives on parking areas), we plan to design an intelligent scenario filtering mechanism to extend the calibration scenarios to the daily driving, which can automatically select multiple appropriate sub-sequences from the entire drive for LiDAR sensor calibration.