Bundle adjustment without iterative structure estimation and its application to navigation

Abstract

This paper describes a new approach for bundle adjustment (BA), which is based on a non-linear optimization of only the camera pose for all the views in a given sequence of images and does not involve iterative structure estimation. If required, structure reconstruction can be performed based on the camera matrices after convergence of the optimization process. Instead of applying the projection equations, the cost function being optimized in the suggested approach is based on the three-view geometry constraints that should be satisfied for any three views with a common overlapping area. Significant reduction in computational complexity is obtained compared to a standard BA, since the number of unknown parameters participating in the iterative optimization is much smaller. The optimization problem is formulated relative to the camera pose of the first view, as commonly used in robotics navigation applications. The proposed method is demonstrated on a publicly available dataset of real images and the optimized camera pose and the recovered structure are compared to the ground truth.