A classification and temporal filtering based system for online extrinsic camera calibration
Jens Westerhoff, Stephanie Lessmann, Mirko Meuter, Anton Kummert · 2016
This paper proposes an online extrinsic camera calibration system able to determine the pitch and roll angle of a moving, forward-looking camera relative to the road surface. Some road surfaces do not show enough texture to allow reliable optical flow calculation and accurate per-frame pitch and roll estimations, respectively. If inaccurate per-frame estimations are regarded in the final calibration, no exact and stable determination of calibration is possible. Therefore, the main contribution of this paper is a classifier system which enables us to distinguish between accurate and inaccurate per-frame angle estimations. To achieve a final calibration only the accurate (classified as positive) per-frame estimations are incorporated in a temporal filter. In our research we found that a fixed amount of positive classified per-frame estimations (a fixed temporal filter length) can be specified which guarantees an exact final calibration. With that, our system is able to autonomously signalizes the moment at which an accurate calibration is achieved while driving. With our developed system, road surfaces with no texture only delay the moment at which a final calibration is achieved but do not affect the exactness of calibration.