Robust three-dimensional scene recovery from monocular image pairs

Abstract

Three-dimensional (3D) scene recovery from two-dimensional (2D) image data is a challenging task. Typical methods applied involve computing 2D image feature correspondences between multiple frames, used in conjunction with camera models and movement to recover 3D feature position. In this work, we present a novel method for 3D scene recovery from monocular video. To provide greater robustness, we propose that rather than searching the 2D image space for feature correspondence, the 3D space in which the recovery is performed is searched directly. However, the search space presented by such a task can be quite large. We, therefore, present a technique to reduce the traversal of this search space. This comprises of enforcing geometric constraints on the recovered 3D data, ensuring a logically consistent 3D scene is reconstructed. We further propose a method (combining sequential image subtraction and region growing) to cope with the problems of 3D scene reconstruction as applied to unconstrained outdoor data. The proposed method is applied to the 3D recovery of outdoor scenes. Comparing with other, more traditional, 3D recovery methods, the proposed method provides more accurate results.