Direct planar odometry with stereo camera

Abstract

Contemporary Visual Odometry (VO) methodologies generally are created upon pointbased approaches to estimate the camera’s pose and a representation of the explored environment. Direct Sparse Odometry (DSO) is the most popular point-based technique of a class of approaches called direct methods. Many works presented improvements in the point-based VO area using DSO characteristics as the fundamentals of these approaches. However, only recently, two new monocular plane-based DSO have been presented. The first approach utilizes a learning-based plane estimator to generate initial plane estimates, which may lead to optimization issues. The second approach restricts plane detection to horizontal and vertical orientations, making it more suitable for structured environments. This thesis presents a stereo plane-based VO technique - Direct Planar Odometry (DPO) - that employs planes as features in a Sliding window optimization framework and utilizes unit dual quaternion as the pose parameter. Our experiments show that the proposed methods achieved comparable results to the Stereo DSO point-based approach.