Machine Learning Based Automatic Extrinsic Calibration of an Onboard Monocular Camera for Driving Assistance Applications on Smart Mobile Devices

Smart mobile devices can be easily transformed into driving assistance tools or traffic monitoring systems. These devices are placed behind the windshield such that the camera is facing forward to observe the traffic. For the visual information to be useful, the camera must be calibrated, and a proper calibration is laborious and difficult to perform for the average user. In this paper, we propose a calibration technique that requires no input from the user and is able to estimate the extrinsic parameters of the camera: yaw, pitch and roll angles and the height of the camera above the road. The calibration algorithm is based on detecting vehicles using CNN based classifiers, and using statistics about their size and position in the image to estimate the extrinsic parameters via Extended Kalman filters.