A Method for Constructing Semantic Navigation Maps in Urban Environments

Abstract

The field of modern autonomous driving and unmanned systems places significant importance on planning and semantic navigation tasks for autonomous vehicles. The successful execution of these tasks relies heavily on mapping and localization technologies. We propose a method to address the difficulty of using trajectory maps of autonomous vehicle cameras in complex environments for planning. The proposed method constructs a topological map that is convenient for subsequent planning and semantic navigation. To enhance the integration of trajectory maps with topological maps, we employ the ORB-SLAM2 algorithm to create the camera's trajectory map, and we extract the keyframe information to obtain frames from the data. These extracted keyframes are then fed sequentially into a junction recognition network that utilizes ResNet18 transfer learning. Then we connect the resulting nodes that are recognized as junctions to generate the camera's topological map. The integration of trajectory and topological maps has been found to be advantageous in improving the planning and navigation abilities of ground robots.