Visual odometry of dense rgb-d images using different optimization methods

Abstract

The goal of this thesis is to accurately estimate the motion of a camera embedded in a robot or a moving object in a static scene using RGB-D images. These images can be provided by a stereo camera or by using a color digital camera as well as one that provides the depth of the scene. Therefore, RGB-D images are the only information acquired by the system from the environment. Thus, the movement of the system is estimated using different consecutive images. The unknown camera motion can be determined by minimizing the intensity error between every two consecutive images. Hence, the challenge of motion estimation is transformed into a non-linear least squares optimization problem, with robot motion being the unknown solution. The solution of such problems typically involves iterative approaches. Exact methods use the linearization of the least square equation to resolve this problem. Alternatively, we can use metaheuristic optimization methods to solve this non-linear equation. Note that the optimal solution will be used to estimate the position of a mobile robot. To evaluate the visual odometry methods, both exact and metaheuristic methods, we apply the root mean square error to an extensive set of images.