Pole-Mapping&Localization

Building a High-definition Semantic Pole-Map and its applications

In most urban and suburban areas, pole structures such as trees or utility poles are ubiquitous. These types of structural landmarks are very useful for the localization of autonomous vehicles given their geometrical locations and measurements from sensors.  In this project, we aim at creating a high-definition navigation map for autonomous vehicles or robots with poles as the dominant localization landmarks. In contrast to the previous pole-based navigation maps or egomotion estimation methods, we exploit the semantics of poles. The main idea is  to segment the scenes into different semantic pieces by a mask-range network, where poles with semantics are obtained in each frame. Simultanously, these semantics are utilized for egomotion estimation as the frontend of a SLAM framework. Loop-closure detection performance can also be significantly improved with these semantics, which incurs a high-definition semantic pole-map. Given this map, we propose a semantic-uncertainty aware Monte Carlo particle-filter localization method for autonomous driving applications. The only sensor for this work is a multiple-channel (16, 32, or 64) LiDAR.