Global Positioning inside Tunnels Using Camera Pose Estimation and Point Clouds

Abstract

This master thesis aims to solve the positioning problem in the absence of an accurate GPS solution. We will show how high density tunnel model in 3D point clouds, together with only a single vehicle-mounted camera, can estimate the vehicle’s position throughout tunnels without the assistance of other positioning systems. We developed a solution by analysing an image sequence where feature detection and image to point cloud backprojection were used to generate an accurate positioning system. The feature detection was performed using the well-known SIFT algorithm and back-projection onto the point cloud is done through repeated closest-neighbour search along a ray we build from the feature’s coordinates. Multiple post-processed GPS runs within the tunnel are used to generate a ground truth which is used as the reference for evaluating the solution. For this end, we developed a new software architecture where the software could ultimately be placed on top of any existing positioning method and then provide an easy comparison between the two. An experimental result is provided to show that accurate positioning can be achieved under the right circumstances, reaching a drift distance of below ten centimeters.