Augmenting Navigation Systems for Near Real-Time Scenarios Using LiDAR




Azam, Muhammad

Journal Title

Journal ISSN

Volume Title



Global Navigation Satellite Systems (GNSS) are critical components of today's intelligent transport applications. They work in conjunction with Road Network Graphs (RNGs), real-world abstractions of transportation infrastructure, to provide services such as location awareness, route-finding, and point-to-point navigation. When the availability or reliability of RNGs becomes an issue, such as in rural environments where data may not be available or in evolving scenarios where real-time data is needed such as disasters, omnipresent navigational systems are unable to cope. In this work, we explore the feasibility of adapting aerially captured Light Detection and Ranging (LiDAR) point clouds for the purposes of both providing a direct means of navigational assistance as well as augmenting existing RNGs. Toward this end, we present a ground segmentation algorithm to identify ground points in a given city-scale point cloud, and a path-planning algorithm building on this segmentation designed to produce plausible paths through an urban area. Our ground segmentation method achieves an average accuracy of 86\% on 36 point cloud dataset tiles from the Sensaturban dataset, performing better on tiles with more points, and completing both segmentation and classification steps in an average of 86 seconds per 1,000,000 points. It also demonstrates effective qualitative performance on a tile from the Vancouver LiDAR dataset. Our proposed path generation algorithm demonstrates an 85\% error reduction in a challenging scenario using only the LiDAR point cloud and its image-analogous gradients as input. We discuss the inexistence of a suitable dataset, presenting a barrier for a large-scale analysis and comparison in this problem. Many current leading techniques in point cloud processing employ some form of learning, furthering the need for such a dataset. We conclude by discussing some design considerations of such a dataset and present directions for future research in this area.