Unmanned Systems Technology 005 | Selex ES Falco UAV | Sense and avoid systems | RCV Engines DF70 | DSEI show report | Fuel cells | CUAV Expo, InterDrone and CUAV Show reports | SLAM

79 out at every point in time, adding to the processing overhead. Local linearisation is only a partial solution, however, as the EKF can still be tripped up by large non-linearities. It is also relatively slow in computing large maps with lots of landmarks because every measurement the vehicle makes affects everything else, so update calculations can take a lot of time. A variant called the Compressed of probable system states (landmark locations and the autonomous system’s position) follow a ‘normal’ bell-curve distribution on a graph, also known as a Gaussian distribution. They also assume that the motion and observation models are linear functions. However, the real world is always much more complex: the curve might look like a bell that has been hit with a hammer and have a very irregular shape, while vehicles might respond to commands or environmental disturbances such as gusts of wind in a non-linear fashion. In these circumstances, assumptions of linear responses and normal distribution can lead to large errors; Kalman filters do not cope well with non-linear functions. Cheating non-linearities The Extended Kalman Filter (EKF) gets around this with a technique called local linearisation, which essentially involves fitting a straight line as closely as possible to an awkwardly shaped distribution curve, passing through the mean point and then proceeding as an ordinary Kalman filter would. This works quite well in practice if most values are fairly close to the mean, but not so well if they don’t, and it also means that new linearising calculations have to be carried Extended Kalman Filter (CEKF) gets around this to an extent by storing information from a small local area, then transferring it to the rest of the map at a lower total processing cost than an EKF. The strangely named Unscented Kalman Filter (UKF) simplifies things by basing its calculations on a carefully chosen set of sample points from the original bell-curve distribution of probable positions for the autonomous system. One technique involves running them through a non-linear function such as one that describes the way the system responds to a motion command, and computing an updated bell-curve distribution to provide a new best estimate of where the autonomous system is. The UKF is much more accurate than the EKF, but is also slow to update when dealing with large numbers of landmarks. The information filter is another variant of the KF, one that avoids linearisation problems by using information vectors and information matrices. It can provide more accurate position estimates, and is considered more stable than the EKF, but it can run into computational trouble during update steps when applied to non-linear systems. Distorted bell curves All varieties of KF, however, can only deal with Gaussian distributions; to deal with more chaotic real-world probability distributions – those that look like badly dented bells – SLAM algorithm developers are turning to particle filters. Sampling techniques are very good at choosing representative samples from Gaussian distributions, so particle filter algorithms take advantage of this by using ‘importance sampling’, a technique that allows sampling from a normal bell curve overlaid on the messier distribution of interest, attaching a higher weighting to samples from areas where the two are very different. Every sample can be seen as a state hypothesis – a possible state the system may be in regarding the unmanned Unmanned Systems Technology | Dec 2015/Jan 2016 Cut off from GPS and potentially all external sources of navigational information, UUVs such as this one are likely to rely heavily on SLAM in the future (Courtesy of NOAA) The best choice of algorithm for a particular environment depends on factors such as the time available for processing