Performance Models for Vision-based SLAM Systems

The navigation of Autonomous Ground and Air Vehicles (AxVs) in urban, potentially GPS denied, environments remains a challenging problem. Simultaneous Localisation and Mapping (SLAM) using inexpensive digital cameras is one attractive technology option for use to support navigation within such environments. However, the performance trade-offs between accuracy of the map or vehicle localisation against the quality of the sensor and other vehicle instruments is not currently well understood. The lack of such understanding makes feasibility assessment and detailed system design difficult. This paper defines performance models to address this need and so to allow the assessment and design of systems incorporating such vision based SLAM components. These models are then used to investigate the system trade-offs in order to provide example ‘data sheets ’ to aid the system designer. Finally, evidence supporting the validity of these models is presented from experiments in simulation based on an EKF based SLAM implementation