Perspective and 360° equirectangular projection image-based stereo-monocular distance estimation with online calibration

Abstract

For the safe autonomous driving with vision-based distance estimation, gaining a wide field of view information to sense the surrounding situation and precise information with details for reliability are essential factors. However, obtaining the two factors simultaneously is difficult, as they relate in inversely due to the limited size of the image sensor. Considering the issue, in this paper, I suggest the new distance estimation system by using two different types of images which are the perspective projection image for the details in the region of interest (RoI) and 360° omni-directional equirectangular projection (ERP) image for the beyond regions. The distance is estimated in stereo-based solution at intersected regions and monocular-based solution at the rest of the regions. However, due to the inaccurate result of the monocular-based solution, I suggest supplementing the inaccuracy by conjugating the stereo-based solution. Besides the distance estimation, as the region of interest (RoI) can be changed due to the user’s attention, the system also includes an online calibration task to consider a rotation of the perspective projection image.