Several renowned techniques exist for Simultaneous Localisation and Mapping (SLAM) – from the classic SLAM that employs the extended Kalman filter to contemporary graph-based algorithms. Our lab has successfully developed a unique SLAM algorithm that takes 3D laser scans as input. What sets our algorithm apart is its application of the Dempster-Shafer theory in place of the Bayesian theory to represent a map’s cell model. In traditional methods, each cell has a specific likelihood of being occupied. However, the Dempster-Shafer theory permits multiple masses to reside in a single cell and provides formulas for updating these masses. This project's core concept is to leverage these masses to discern patterns in the map. Such patterns can manifest as distinctive object borders or areas that correspond to entire objects within the environment.
With monitoring point cloud profiles, we can use the Maximum Covariance Unfolding method. This technique enables the processing of a reduced dimension of the resultant profile, while preserving its relationship with the explanatory variables. These variables depict the acquired pattern based on how certain features deviate from a reference. Among other factors, the explanatory variable can represent the pattern's angle of deviation along a specific axis. We are currently exploring the use of Maximum Covariance Unfolding to swiftly identify 90-degree angles, which are prevalent in indoor environments, to assist in robot localization within such spaces.