Extrinsic Camera Calibration Using Vehicular Interior Features - A Domain-Agnostic and Targetless Pipeline

Abstract

Advanced driver assistance and safety systems increasingly rely on interior sensing systems, such as Occupant Monitoring Systems (OMS) and Driver Monitoring Sys-tems (DMS). These systems depend on accurately calibrated interior cameras to provide reliable detection of behavior and positioning of the driver and occupants. Traditional camera calibration often relies on dedicated targets such as checker-boards, but such methods can be difficult to integrate into large-scale production environments. This thesis investigates targetless extrinsic calibration of an OMS camera by esti-mating its six degree of freedom pose from a single image of the car interior. Instead of using calibration targets, the proposed approach uses interior features as reference points. Two methods are evaluated: a relative method using a nominal reference im-age, and an absolute method using triangulated 3D points from known established views. The results show that extrinsic calibration using interior features is possible, with especially strong rotation estimation. Both methods presented here achieve errors as low as 0.09 degrees on synthetic data, while translation proved more challenging within the small mounting tolerances. The Absolute Method performed best for translation on synthetic data, reducing the error from an average of 5 mm to 1 mm, whereas the Relative Method showed no reduction in error. The Relative Method did however show better performance on limited real vehicle data by consistently reducing the reprojection error compared with the uncalibrated baseline. Overall, the thesis demonstrates the potential of targetless calibration for scalable OMS camera deployment. Although it is demonstrated on car interiors, the proposed pipeline is domain-agnostic and may be applicable to other environments containing repeatable visual features.