Rao-Blackwellized Particle Filter Approach to Monocular vSLAM With a Modified Initialization Scheme

Abstract

This paper presents a Rao-Blackwellized particle filter (RBPF) approach with a modified undelayed initialization scheme to solve the 3D visual SLAM problem (vSLAM) using a single camera. In the proposed method, landmarks are initialized using the inverse depth of the landmarks rather than the traditional use of their depths. In this scheme, there is no need to distinguish between partially and fully initialized landmarks. Once the landmarks are properly initialized, the RBPF enhances the estimation of the robot path and landmark location using bearing-only information obtained from a camera. The results of numerical simulations and experiments with a video clip have been included in this paper to verify the performance of the proposed approach.