Mapping of Parking Areas using Radar Sensors with a Cluster-based Landmark Extraction Algorithm and an Extended Kalman Filter

Abstract

This thesis investigates the possibility of building maps of partially filled parking areas, accurate enough to find empty spaces, using Radio Detection and Ranging (radar) sensors. Existing techniques for solving feature-based Simultaneous Localization and Mapping (SLAM) will be used as basis for map building, but refinements are needed in order to handle noise from radar sensors. A new landmark extraction algorithm is developed for finding lines and corners within radar data. The algorithm first cluster detections that belong to the same car using a single-linkage clustering, then lines and corners are found within each cluster by a line segmentation algorithm. The landmarks are used in two different SLAM approaches. The first is a standard SLAM approach using an Extended Kalman Filter (EKF) in combination with single lines as landmarks. The second includes an additional step using an Extended Information Filter (EIF) to maintain correlations between features within more complex landmarks, such as lines and corners of the same car. Precision and correctness of the algorithms are evaluated in real world scenarios using Light Detection and Ranging (lidar) data in a line by line comparison. Results show that EKF-SLAM maps are noisy, but have most lines located close by cars. It is possible to detect free spaces within the maps, even though noise is present and some lines are too short. Including the EIF correlation step shows promising results for creating less noisy maps, however the landmark extraction limits its performance in dense parking areas. Both approaches can create maps where it is possible to locate available parking spaces.