Recasting SLAM - Towards improving efficiency and platform independency

Abstract

This paper provides an alternative solution to solving SLAM's computational complexity in Inertial Navigation System (INS) application, not from the perspective of map management techniques, but by focusing on the filter's structure and model, and recasting the SLAM algorithm into what is known as an "indirect" implementation. In doing so we provide a navigation structure which is computationally efficient in even for highly non-linear, highly dynamic systems. The problem is solved by separating the SLAM filter from the main navigation loop and, instead of estimating the states of the vehicle and landmarks, the filter estimates the errors in these states. This is accomplished by perturbing the dynamic equations which govern the platform and map models, and hence linearising an otherwise highly non-linear system. The error behaviour of INS is well known and can be predicted precisely using the linearised model. The result is a SLAM linearized error model which provides four main benefits: 1) since the model is linearised, the estimation filter itself is linear during sample interval, providing both significant advantages in computation and in filter tuning; 2) the error model represents the error dynamics, which drift slowly with time, hence the sampling rate required for the prediction cycle of the filter is significantly lower; 3) since the navigation loop and map are decoupled from the time-consuming filter structure, the navigation loop can provide the navigation outputs within fixed deadline without disturbed from the filter; and 4) as the error model is in piecewisely linear form, the observability of SLAM system can be directly analysed. Furthermore, in this paper the navigation structure makes use of an INS as the driver for the platform model, thus providing a navigation solution which is totally separated and independent of the vehicle implemented (however the structure implemented here can also be used when given a vehicle kinematic representation cast into an error model form). Results from the implementation of indirect SLAM on an airborne vehicle illustrates that the filter structure can statistically estimate the errors and provide a navigation solution which is comparable to that of the direct SLAM structure however with significantly less computational cost, and without the need for a vehicle model.