Self-Supervised Online Camera Calibration for Autonomous Driving Applications

Camera calibration is a critical step for accurate perception in autonomous vehicles. However, traditional calibration methods are often laborious and time-consuming, and require specialised data collection and careful tuning. This process must be repeated whenever the parameters of the camera change, which can be a frequent occurrence due to vibration, impacts or extreme temperatures. This paper proposes a novel deep learning framework for self-supervised camera calibration. The framework learns the intrinsic and extrinsic camera parameters from unlabeled driving data. This allows the framework to calibrate the camera in real time, without the need for any physical targets or special planar surfaces. This could cut the cost of re-calibration for both the manufacturer and consumer as the vehicle can re-calibrate itself over the lifetime of the vehicle, which also improves the accuracy and safety of the ADAS systems on the car