Three dimensional scene reconstruction from passive imagery

This paper describes a new approach, developed under the auspices of the Electro Magnetic Remote Sensing Defence Technology Centre, for reconstruction of three dimensional scene content using passive video. This gradient based scheme uses the parallax in an image sequence to estimate the distance from sensor to scene. Intensity measurements from a passive camera are combined with a rigid body model of scene content to infer both the motion of the camera relative to the scene, and the 3D scene structure. The latter is stored as a ‘depth map’, describing the distance from sensor to scene for every pixel. Inference of the camera motion is via an efficient Bayesian scheme, HINTS, which estimates the incremental change in the camera position and pose, between successive images. For the inference of both the camera motion and the scene depth map, a recursive framework is adopted which seeks to minimise an error defined using the intensity in the measured images. The resulting scene depth estimate is illustrated using airborne imagery.