Image-based models with applications in mobile robotics

• Authors: Hong Zhang;Dana Cobzas
• Venue: Image-based models with applications in mobile robotics
• Year: 2004

In computer vision and graphics as well as in mobile robotics one is often interested in capturing 3D real world scenes. In the former, models are captured for such purposes as photorealistic rendering. In the latter, models are captured to generate navigation maps for robot control. The objective in both fields is similar: How can the modeling of existing scenes be achieved? This thesis investigates alternative ways of capturing the geometry and appearance of an indoor environment by using image-based modeling. Calibrated and uncalibrated methods are investigated and contrasted. The resulting models are validated by utilizing them in two robotics tasks. We first use the model as a navigation map to localize the pose of a robot and to track its motion based on images from a camera fitted on the robot. Second, we generate synthetic images by reprojecting and texturing the captured model given a desired camera pose. The first approach is based on a panoramic image mosaic augmented with depth information and is built using calibrated cameras and range sensors (a trinocular device and a laser range-finder). Several methods for registering camera and range sensors were developed and compared. The model is segmented into planar pieces that can be reprojected in new positions. The second approach uses an uncalibrated camera that samples a scene. By maintaining visual tracking of corresponding feature points, the geometry of the scene is reconstructed using stratified structure from motion. The geometric model is then bundle-adjusted and reprojected into the original images to acquire surface appearance. Surface appearance is represented not using a single traditional texture, but by acquiring a basis that captures the view dependency of the surface.