Adaptive sequential data association in SLAM

The nearest neighbor data association algorithm based on Mahalanobis distance has been widely used in SLAM. It is usually assumed that the Mahalanobis distance between the estimation and the observation of a landmark follows a Chi-square distribution. So the data association thresholds will be determined fixedly according to the given confidence levels. It is proved, however, with the ground truth SLAM simulations, that the Mahalanobis distance does not always follow a Chi-square distribution, but it is affected by the distribution of landmarks, process noise, measurement noise and the magnitude of control inputs. Further, a virtual association based data association method was proposed. It could produce adaptive thresholds by using all the observations and the pose estimation of a mobile robot. A sequential data association approach was adopted to prevent augmentation delay and missing landmarks. Simulation results show that the adaptive sequential data association algorithm adapts well to the change of environments and noises, and it can also avoid the false landmarks and reduce the rate of missed observations.