Efficient 3D reconstruction using monocular vision

For airborne vehicle flying at low altitudes, recovery of depth map in front of aircraft is very important to avoid obstacle. This paper describes a novel method for scene reconstruction using a sequence of images from a passive sensor (TV, FLIR) and the information from inertial navigation system. We model 3D scenes in the field-of-view as a collection of approximately planar layers that corresponds to the background and obstacles respectively. Therefore the central problems of reconstruction are the depths calculation of every layer and the assignment of pixels to layers. Then these parameters are estimated within Bayesian framework. A graph cut algorithm is used to efficiently find the approximation. Experimental data demonstrates the effectiveness of our approach.