Camera Calibration for CNN Based Generic Obstacle Detection

The Convolutional Neural Networks (CNNs) have made monocular image processing a powerful obstacle detector, but in order to transform these results into 3D data robust automatic calibration is needed. This paper proposes an unassisted camera calibration algorithm, based on analyzing image sequences acquired from naturalistic driving. The focal distance is computed based on the mean lateral displacement of similar features in consecutive frames, compared with the yaw rate of the vehicle. The height and pitch angle are computed based on the distribution of the lane width values on the image lines, assuming an average lane width is known. The lane markings are detected using the edges on the road, already segmented using the CNN. The yaw angle is computed using the vanishing point (VP) detection, which is performed using the direction of the road gradients. The pitch angle value is dynamically corrected using the VP, and using comparisons between the past frame and the current frame, under multiple correction hypotheses.