Stereo Vision Specific Observation Model for EKF-Based SLAM

Abstract

This work addresses a new probabilistic observation model for a stereo simultaneous localization and mapping (SLAM) system within the standard Extended-Kalman filter (EKF) framework. The observation modal was derived by using the inverse depth parameterization as the landmark modal, and contributes to both bearing and range information into the EKF estimation. In this way the inherently non-linear problem cause by the projection equations is resolved and real depth uncertainty distribution of landmarks features can be accurately estimated. The system was demonstrated with real-world outdoor data. Analysis results show landmark feature depth estimation is more stable and the uncertainty noise converges faster than the traditional approach.