Tham khảo tài liệu 'mobile robots perception & navigation part 9', kỹ thuật - công nghệ, cơ khí - chế tạo máy phục vụ nhu cầu học tập, nghiên cứu và làm việc hiệu quả | Robust Autonomous Navigation and World Representation in Outdoor Environments 311 indication that the filter cannot continue working assuming a mono-modal probability density distribution. At this time we have the CEKF estimated mean and deviation of the states representing the vehicle pose and landmark positions. With the currently estimated map a decorrelated map is built using a coordinate transform and a decorrelation procedure Guivant Nebot 2002 . A particle filter Gordon et al. 1993 is initialised using the available statistics and is then used to resolve the position of the vehicle as a localisation problem. Once the multi-hypothesis problem is solved the CEKF is restarted with the states values back propagated to the time when the data association problem was detected. Then the CEKF resumes operation until a new potential data association problem is detected. There are several important implementation issues that need to be taken into account to maximise the performance of the hybrid architecture proposed. The solutions they need to consider are the uncertainties in vehicle map and sensor to maximise the number of particles in the most likely position of the vehicle. The SLAM algorithm builds a map while the vehicle explores a new area. The map states will be in most cases highly correlated in a local area. In order to use the particle filter to solve the localisation problem a two dimensional map probability density distribution needs to be synthesised from an originally strongly correlated n dimension map. The decorrelation procedure is implemented in two steps. The map originally represented in global coordinates is now represented in a local frame defined by the states of two beacons that are highly correlated to all the local landmarks. The other local landmarks are then referenced to this new base. A conservative bound matrix can be obtained as a diagonal matrix with bigger diagonal components and deleting the cross-correlation terms Guivant Nebot .
