Sparse-Map: automatic topological map creation via unsupervised learning techniques

摘要

Most robots use 2D occupancy grid maps for navigation, localization, and path-planning. This model is flexible and allows to represent any geometrical shape with finite accuracy. However, this dense representation imposes high memory requirements and does not generalize well to 3D environments. We present a task-based map compression technique useful for path-planning and navigation in indoor environments for service robots where, from a point cloud of 3D map features, we calculate a number of clusters based on their spatial position and generate a sparse 3D representation of the environment. Moreover, we propose several metrics to assess the quality and performance of a map representation, and we tested our proposal using a series of point-cloud benchmarks and clustering techniques where our method has a comparable performance using a fraction of the memory footprint than the baselines. Finally, we have released our system as a Robot Operating System (ROS) based open source library.