engleski

Simultaneous mobile robot localization and three-dimensional modeling of unknown complex environments in real time

The two key prerequisites for autonomous navigation of any mobile robot are its capability to infer (1) what the environment around it looks like and (2) where it is in the environment. To infer where it is a robot localization algorithm is needed, while to infer what the environment looks like a mapping algorithm is needed. The difficulties with estimating the environment map and the robot location arise from the fact that they are strongly interrelated. For this reason the present thesis focuses on development of the Simultaneous Localization and Mapping (SLAM) algorithms which allow mobile robots to build the map of completely unknown environments, and at the same time estimate their location within. The main problems of today’s state of the art SLAM algorithms are: (i) to find appropriate map representation, (ii), computation speed, (iii) to achieve long-term autonomy, and (iv) to distribute the SLAM tasks over several robots. In this thesis all four problems are addressed. For the map representation, the algorithm is presented which allows modeling of 3D complex environments using as small as possible number of planar surface segments. Moreover, the algorithm also allows for fast update of the entire map when the SLAM optimization is completed. Computation speed is addressed through the derivation of the Exactly Sparse Delayed State Filter (ESDSF) on Lie groups (LG- ESDSF). LG-ESDSF represents the states on Lie groups, and performs filtering equations in the pertaining Lie algebra, which allows it to respect the geometry of the state space. Because of this LG-ESDSF achieves accuracy comparable to the state-of-the-art SLAM algorithms while maintaining high computation speeds. A solution for long-term SLAM is also presented in the context of LG-ESDSF. It detects which states have negligible information gain and removes them from the state space, but at the same time preserves the sparsity of the information matrix, thus allowing long-term real-time performance possible. Finally, the distribution of the SLAM tasks over several agents is solved by the introduction of the cooperative SLAM algorithm. The proposed algorithm allows each agent to compute only simple computing tasks (trajectory optimization in our case), and therefore they can have low computational power and still work in real-time, while the central server performs complex computing tasks (global map building in our case). Moreover, server collects information from all the robots and sends the update information based on data from one agent to the other agents. Agents then use this information to further increase their trajectory accuracy.

simultaneous localization and mapping, filtering SLAM, graph-based SLAM, exactly sparse delayed state filter, 3D planar map, Lie groups, cooperative SLAM, long-term

nije evidentirano