A Visual SLAM Solution Based on High Level Geometry Knowledge and Kalman Filtering

In this paper, two new methods are proposed for robotic simultaneous localization and map building (SLAM), namely high level geometric knowledge constraint and newly acquired feature initialization. These methods are implemented within classic extended Kalman filter (EKF) framework. Novelties lie in two aspects. First, high level geometric information, such as common geometric primitives (e.g. lines and triangles) constructed by observed feature points, is incorporated to EKF to enhance the robustness and resistance to noise. Second, a visual measurement approach, multiple view geometry (MVG), is employed for new feature initialization that is considered as a key factor affecting the lower bound error in robotic mapping. Simulations are performed, which can be deemed as concrete verifications and extensions to previous results reported by other researchers. The numerical results show great potentials