Doppler radar odometry for localization in difficult underground environment

Sammanfattning: Accurate and efficient localization is a fundamental requirement for autonomous operation of robots, especially in areas that deny global navigation services. Localization is even more challenging in environments that present visual and geographic difficulties. This not only includes environmental aspects like darkness, fog and dust but also geometrically monotone areas. The solution that the Center for Applied Autonomous Sensor Systems at the Örebro university decided to develop is therefore a prototype of radar-only localization and mapping (SLAM) system. The radar modality is less susceptible to the environmental factors when compared to, for example, a lidar. Our goal is to support this effort by creating an odometry module that uses radar and inertial data to provide the localization for this SLAM prototype. This radar-inertial-odometry (RIO) takes radar point clouds and inertial gyroscopic data to output an odometry message usable by other components in the robot operating system (ROS). The module has been tested on two datasets representing areas typical for deployment, one consisting of underground tunnels and the other one being an outside forest environment. The dataset has been processed by two different mappers where the lidar has been used as the basic modality. This choice allows us to evaluate the odometry module in a more practical way. The final results are promising, the underground localization closely adheres to reality. The forest dataset is more challenging although it still resembles the ground-truth position in the horizontal dimension. The module's biggest shortcoming is a noticeable drift problem in the vertical z-dimension , for which we propose a constraint that limits this drift.