Dense 3D Structure and Motion Estimation as an Aid for Robot Navigation

Three-dimensional scene reconstruction is an important tool in many applications varying from computer graphics to mobile robot navigation. In this paper, we focus on the robotics application, where the goal is to estimate the 3D rigid motion of a mobile robot and to reconstruct a dense three-dimensional scene representation. The reconstruction problem can be subdivided into a number of subproblems. First, the egomotion has to be estimated. For this, the camera (or robot) motion parameters are iteratively estimated by reconstruction of the epipolar geometry. Secondly, a dense depth map is calculated by fusing sparse depth information from point features and dense motion information from the optical flow in a variational framework. This depth map corresponds to a point cloud in 3D space, which can then be converted into a model to extract information for the robot navigation algorithm. Here, we present an integrated approach for the structure and egomotion estimation problem.