Automatic LiDAR-camera calibration: Extrinsic calibration for a LiDAR-camera pair using structure from motion and stochastic optimization

Abstract

This thesis presents an approach to automatically and simultaneously perform extrinsic calibration of a LiDAR and a camera. Nowadays, a multitude of sensors are used in a majority of vehicles. Having correctly calibrated sensors is essential for attaining accurate data to use in various sensor dependent applications. Today’s LiDAR-camera calibration methods are often performed manually or require externally introduced calibration objects. However, the method proposed in this thesis is only dependent on 3D LiDAR point clouds and camera images. The method consists of two major parts. Firstly, the camera images were converted to 3D point clouds using a structure from motion pipeline, ensuring that the data from both sensors were comparable. Secondly, a genetic algorithm with an objective function based upon a 3D voxel grid filter was used to iteratively compare the overlap of the point clouds until convergence. The method proved to be successful in creating 3D point clouds from camera images and accurately estimating the rotational parameters for both sensors. However, it was not as robust and accurate as anticipated when estimating the sensor positions.