A robust feature extraction method and semantic data association for 6D SLAM

Simultaneous Localization and Map (SLAM) building problem has vital importance for fully autonomous systems in partly or fully Global Navigation Satellite Systems (GNSS) denied environments. For LiDAR systems, there are two different ways to deal with SLAM problems. They are graph representation based SLAM methods depending on mostly scan matching algorithms and probabilistic SLAM methods based on feature extraction and landmark assignment. Since the scan matching methods solves the registration problem of two successive scans, they only provide the knowledge of relative movement of the mobile robot. However, feature based probabilistic SLAM methods update the whole map and robot pose when a new observation is made. However, the difficulty in feature based SLAM method is the problem of extracting reliable features. While in 2D mostly line, corner, and single points are used as features, in 3D the favorite feature is the plane. The plane extraction algorithms works well in indoor and structured environments, but they have problems in outdoor and especially in complex environments. To be able to merge two planes, two conditions must be satisfied, which are based on the plane orientation and translation in normal direction. However, this is not sufficient in outdoor case since there might be two planes segments that satisfy these conditions but actually far away from each other. To deal with this problem, in this paper, in addition to two conditions, we propose a new constraint while merging two planes. This constraint is based on pooled covariance matrix and Mahalanobis distance function. The second contribution is to use Principal Component Analysis (PCA) in the projection phase which is applied before finding the convex hull of the plane segment. Moreover, the data association problem, which is very problematic in point feature based SLAM, is solved easily by using the semantic data obtained from the plane detection method. The experimental results show that these features can be used in feature based SLAM successfully.