Stereo Reconstruction for Visual Navigation

Visual Navigation is the key to enabling useful mobile robots to work autonomously in unmapped dynamic environments. It involves positioning a robot by tracking the world as it moves past a camera. This paper describes techniques we are developing to improve the accuracy of transformations calculated by matching sets of points between two views. This tells us how far the robot has moved. In the best contemporary Visual Navigation algorithms an approximate estimate of the transformation is improved using the Bundle Adjustment algorithm to incorporate all information from point correspondences tracked over multiple views. This iterative optimisation procedure is computationally intensive and prone to convergence to false minima however, so it is desirable to start as close to the true solution as possible. We compare existing algorithms used to extract a transformation for this initial solution and adapt them to provide improved performance when tracking 3d points from stereo cameras that have uncertain depth.