Online stereovision calibration using on-road markings

Stereo vision helps autonomous vehicles a lot to complete the tasks such as obstacle avoidance, positioning and 3D reconstruction. However, the 6-DOF transform between the camera pair, which is necessary for 3D structure recovery, will not stay constant during the long-time working. In automobile applications, conventional calibration approaches that use man-made patterns are inconvenient and the existing self-calibration techniques cannot produce enough accurate calibration results. In this work, we propose a self-calibration method to recalibrate stereo rig automatically as well as precisely by utilizing onroad markings. First, the system detects crosswalks and extracts their corners in the two images respectively. Next, the system matches the extracted corners, both between the image pair and between images and the corner map. Finally, with matched corner pairs, the stereo transformation of the stereo rig will be estimated. Experiments show that the proposed approach could extract and match crosswalks´ corners accurately, and recalibrate sterevision system in terms of higher accuracy and robustness.