This publication shows a practical approach to evaluate the metric accuracy of the complete image processing chain from calibration over rectification up to stereo vision using vehicle mounted sensors. A Velodyne LASER scanner is used for ground truth. It can be replaced by a manual LASER ranger if the accuracy of the lateral position is of no interest. We calibrate our cameras using multiple views of planar chequer boards of known geometry and unknown position. The extracted corner positions are used in a Bundle Adjustment method to obtain the intrinsic and extrinsic parameters of the cameras, including distortion and relative transform. The transform between scanner and camera coordinate systems is obtained either in a separate step or together with the intrinsic parameters of the camera. Ground truth is obtained from a single corner of an 1 m2 chequer board. The chequer board corner is selected manually and located with same algorithm as in the calibration method. We also manually select the plane of the chequer board in the scanner data and compare the scanner distance to the distance obtained from the cameras by Forward Section. To avoid synchronisation problems, all tests are done with a stationary vehicle. A video of a person placing the chequer board in all positions reachable from the ground is used to verify the typical rules of thumb for chequer board calibrations using a Monte-Carlo simulation. We generated about 1 million different calibration inputs, consisting of 20 image pairs each and computed the calibrations for each set of images. Using the ground truth data above, we found the best, worst, and median calibration based on reconstruction error. The rules of thumb could be verified: Board positions in different depths, different angles, and in the image corners are the most helpful for a small reconstruction error. Our chequer board can be detected at distances of up to 20 m. For the driver assistance system in our car, scene reconstructions up to 50 m are relevant. As a larger chequer board is impractical, we need to quantify the systematic error of the calibration due to extrapolation. The large chequer board corner can be seen in sufficient resolution in distances of more than 95 m. Here, the vertical resolution of the scanner becomes the limiting factor. The processing chain of our driver assistance system rectifies the input image so that pixel rows correspond to epipolar lines by image warping using bilinear interpolation. This might also influence the scene reconstruction. We rectify all images for which we have ground truth and repeat the localisation using Forward Section. The last step is to reconstruct the scene using the stereo vision method from the driver assistance system and compare its results to the ground truth.
Evaluating the accuracy of camera calibration for driver assistance systems
2011
9 Seiten, 9 Bilder, 10 Quellen
Aufsatz (Konferenz)
Englisch