Row Feature Tracking for Improved Localization of Agricultural Robots

This paper presents an approach to utilize the crop rows on agricultural fields for a GNSS-independent navigation. Therefore, row line segments and row end points are defined as row features. They are detected based on LiDAR sensor information and then associated with an a priori derived row feature map, which can be constructed during an initial mapping run. Based on the association of the line segments, the pose of the robot is corrected lateral to the row direction. Additionally, the end point associations allow a longitudinal correction at the end of the row. This corrects the localization drift of relative localization measurements from IMU, wheel odometry, or visual odometry and replaces the need of absolute localization from Real Time Kinematic (RTK) GNSS. The proposed framework is validated in a simulation within ROS and Gazebo.