A pose pruning driven solution to pose feature GraphSLAM
To build consistent feature based map for the environment, GraphSLAM forms the graph using the collected information, with poses of robot and features being nodes while the odometry and observations being binary edges (edge links to two nodes). As the number of kept nodes grows unboundedly while robot moves, this method will become intractable for long duration operation. In this paper, we propose a pose pruning driven solution for pose feature SLAM by relating the size of graph to the size of
... aph to the size of map instead of the length of trajectory. It consists of two steps: (1) An online pose pruning algorithm that can select a pose to be pruned based on the contribution of the pose. Different from conventional methods considering the spatial distance between poses, the contribution is based on the feature observations of poses, taking mapping into consideration. (2) An edge generation algorithm that can build new consistent binary edges from n-nary edge (edge links to n nodes) induced by marginalizing the pruned pose. The type of new edges remains invariant (i.e. they are either odometry or pose to feature observations), so no extra change is required to be made on the GraphSLAM optimizer, making the proposed solution modular. In the experiment, we first employ this system on simulation datasets to show how it works. Then the large scale datasets: DLR, Victoria Park and CityTrees10000 are used to evaluate its performance.