Output details
15 - General Engineering
University of Oxford
An O(N(2)) square root unscented Kalman Filter for visual simultaneous localization and mapping.
In 2007 the state of the art in real-time visual mapping was Davison's monoSLAM. During EPSRC project GR/S97774 “Wearable Visual Computing” we studied two drawbacks inherent in monoSLAM's use of the extended Kalman Filter, viz computational complexity and vulnerability to systemic non-linearity. The Unscented Kalman Filter handled non-linearity better, but its standard form exhibited cubic complexity in the number of scene points, higher than that of the EKF. By analysis of the underlying linear algebra, this paper showed that the UKF could also achieve quadratic complexity. The paper is chosen for its rigorous contribution to vision theory.