By Daniela E. Sánchez, Harvey C. Gómez and Thomas Pany, Institute of Space Technology and Space Applications (ISTA)
This paper presents how our system, consisting of a GNSS receiver antenna, an inertial measurement unit (IMU) and a lidar, is used to obtain high-precision maps through the geo-referencing of lidar point clouds. An accuracy assessment of the system is conducted, which also gives us insights on the quality of lidar range measurements for autonomous driving applications.
Figure 1. Rendering of three georeferenced lidar point clouds on Google Earth. (Image: Authors)
The assessment is done by geo-referencing the obtained point clouds of extracted buildings and comparing them against a supporting measuring system like a total station. The building extraction is done by performing an approximation of the mathematical model of a plane to the facades that composes the building in both, the lidar and the supporting measurement system data.
The paper also indicates the proposed pose determination method of a mobile agent using lidar data. Thanks to the advantages of active, 3D sensors, diverse objects in the environment can be detected as individual point sets, or clusters. Each of the segmented objects can be used as a landmark to figure how the agent is located with respect to those structural elements. The algorithm is capable of detecting the clusters in one point cloud, and finding the most alike point set on a subsequent scan. This is achieved by comparing global descriptors for point cloud data.
The Ensemble of Shape Functions (ESF) is selected as the cluster descriptor. The cluster matching is performed by comparing the clusters one-to-one, calculating the minimum Chi-squared distance among their descriptors. The smaller this distance, the greater the probability of being the same cluster in distinct epochs.
Figure 2. Direct geo-referencing of lidar data at different times. (Image: Authors)
The resultant cluster correspondences for the whole point cloud allow finding the rigid transformation between the point clouds. An initial coarse alignment among the clouds based on the centroids of each matched cluster was performed, followed by a fine alignment in order to reduce errors by the use of the Iterative Closest Point (ICP) algorithm. This approach is valid for urban environments, or for those where many objects can be segmented as clusters.
Finally, a practical case is described in order to show how we plan to use the outcome of the highly precise geo-referenced point clouds and the pose estimation method using lidar.