CSSE Seminar Series (CSSESS)

Quick links: Past seminars,  future seminars,  CSSESS Home

Robust long-term positioning for autonomous mobile robots is essential for many applications. In many environments this task is challenging, as errors accumulate in the robot's position estimate over time. The robot must also build a map so that these errors can be corrected when mapped regions are re-visited; this is known as Simultaneous Localisation and Mapping, or SLAM.

One attractive sensor for SLAM is a digital camera, which captures images that can be used to recognise where the robot is, and to incrementally position the robot as it moves. However, many contemporary schemes for SLAM using only a single camera suffer complete failure in dynamic or featureless environments, or during erratic camera motion. An additional problem, known as scale drift, is that cameras do not directly measure the scale of the environment, and errors in relative scale accumulate over time, introducing errors into the robot's speed and position estimates.

This presentation describes BoWSLAM, a single camera SLAM scheme designed to address these difficulties. BoWSLAM uses a Bag-of-Words scheme to recognise places the robot has visited previously, without any prior knowledge of its environment.

The BoW scheme is also used to select the best set of frames to reconstruct the robot's position from, and is used to efficiently find wide-baseline correspondences between many pairs of frames. Relative positions computed from these correspondences are used to build a pose graph; redundancy in this pose graph enables the selection of only reliable position estimates in the presence of errors. Finally, as the robot explores, objects in the world are recognised and measured.

These measurements enable scale drift to be corrected. BoWSLAM is demonstrated mapping a 25 minute 2.5km trajectory through a challenging and dynamic outdoor environment in real-time, and without any other sensor input; considerably further than previous single camera SLAM schemes.

Tom Botterill conducted his PhD research with the Geospatial Research Centre, and was supervised by Steven Mills from the GRC and Richard Green in computer science. His research interests include the use of computer vision for robot positioning and pedestrian navigation, robust structure-from-motion estimation, and multi-view 3d reconstruction. He now works as a Research Fellow on the MSI-funded "Vision-based automated pruning" project.

Quick links: Past seminars,  future seminars,  CSSESS Home