Georg Nebehay: Evaluation of a self-calibration method for smart video cameras

This report presents an evaluation of a novel edge-based self-calibration method for single cameras in urban imagery. The method uses a combination of RANSAC (RANdom SAmple Consensus) and EM (Expectation-Maximization) for achieving robust calibration results. For evaluation purposes, we made use of a newly available York Urban ground truth database and show that computation time can be reduced by up to 80% by a number of measures. A novel C implementation was created for this report. The evaluation shows that in half of the dataset our calibration method achieves 5% relative error or less in the focal length at a mean computation time per image of 14.45s on a standard PC.