Improved Visual SLAM Using Semantic Segmentation and Layout Estimation

Abstract

The technological advances in computational systems have enabled very complex computer vision and machine learning approaches to perform efficiently and accurately. These new approaches can be considered a new set of tools to reshape the visual SLAM solutions. We present an investigation of the latest neuroscientific research that explains how the human brain can accurately navigate and map unknown environments. The accuracy suggests that human navigation is not affected by traditional visual odometry drifts resulting from tracking visual features. It utilises the geometrical structures of the surrounding objects within the navigated space. The identified objects and space geometrical shapes anchor the estimated space representation and mitigate the overall drift. Inspired by the human brain’s navigation techniques, this paper presents our efforts to incorporate two machine learning techniques into a VSLAM solution: semantic segmentation and layout estimation to imitate human abilities to map new environments. The proposed system benefits from the geometrical relations between the corner points of the cuboid environments to improve the accuracy of trajectory estimation. Moreover, the implemented SLAM solution semantically groups the map points and then tracks each group independently to limit the system drift. The implemented solution yielded higher trajectory accuracy and immunity to large pure rotations.