Reducing the along-track error of point cloud registration through multi-sensor fusion: a case study for machine steering at a composting site

In precision agriculture, vision-based systems are often used for machine steering. When a 3D point cloud of the site exists and the machine is equipped with a stereo camera, the pose of the machine can be estimated through Normal Distributions Transform (NDT). However, in homogeneous environments with elongated features such as windrows at a composting site, registration is difficult and large along-track errors occur. This is disadvantageous, especially when a compost turner should be steered automatically along the windrows.
This paper aims at reducing the along-track error of the point cloud registration through multi-sensor fusion using GNSS-RTK, odometry and a MEMS IMU in an error state Kalman filter. At first, the point cloud registration algorithm is coupled with a dynamic model which uses odometer data to propagate the vehicle state. Secondly, GNSS-RTK and IMU measurements are added. To evaluate both proposed approaches, tests are carried out at a real composting site where sensor data are recorded and later replayed for real-time simulation. The resulting position estimates are compared to a highly precise ground truth to assess how the along-track error improves. The results show that adding odometry helps to reduce the maximum along-track error. However, when NDT fails and only odometry is used to propagate the state vector, the trajectory is affected by drift. By adding GNSS-RTK and IMU observations, the along-track, cross-track and heading errors are significantly reduced. The final multi-sensor solution is sufficiently accurate for machine steering.