Online Mapping and Perception Algorithms for Multi-robot Teams Operating in Urban Environments.

Abstract

This thesis investigates some of the sensing and perception challenges faced by multi-robot teams equipped with LIDAR and camera sensors. Multi-robot teams are ideal for deployment in large, real-world environments due to their ability to parallelize exploration, reconnaissance or mapping tasks. However, such domains also impose additional requirements, including the need for a) online algorithms (to eliminate stopping and waiting for processing to finish before proceeding) and b) scalability (to handle data from many robots distributed over a large area). These general requirements give rise to specific algorithmic challenges, including 1) online maintenance of large, coherent maps covering the explored area, 2) online estimation of communication properties in the presence of buildings and other interfering structure, and 3) online fusion and segmentation of multiple sensors to aid in object detection.

The contribution of this thesis is the introduction of novel approaches that leverage grid-maps and sparse multi-variate gaussian inference to augment the capability of multi-robot teams operating in urban, indoor-outdoor environments by improving the state of the art of map rasterization, signal strength prediction, colored point cloud segmentation, and reliable camera calibration.

In particular, we introduce a map rasterization technique for large LIDAR-based occupancy grids that makes online updates possible when data is arriving from many robots at once. We also introduce new online techniques for robots to predict the signal strength to their teammates by combining LIDAR measurements with signal strength measurements from their radios. Processing fused LIDAR+camera point clouds is also important for many object-detection pipelines. We demonstrate a near linear-time online segmentation algorithm to this domain. However, maintaining the calibration of a fleet of 14 robots made this approach difficult to employ in practice. Therefore we introduced a robust and repeatable camera calibration process that grounds the camera model uncertainty in pixel error, allowing the system to guide novices and experts alike to reliably produce accurate calibrations.