Performance Evaluation on Map-Based NDT Scan Matching Localization Using Simulated Occlusion Datasets

Abstract

This letter presents a performance evaluation on the conventional normal distribution transform (NDT) map-based scan matching under the presence of occlusion. The LiDAR map-based localization method enables centimeter level accuracy positioning; however, the state-of-the-art algorithms do not achieve the same performance when excessive unexpected objects, such as pedestrians or dynamic vehicles, occlude the field of view of the LiDAR. Although the NDT scan matching is able to cope with slight geometrical change of environment, the presence of unexpected objects still induces matching error due to the discrepancy created between the real-time scan and the prebuild map. In this letter, we manually place bounding boxes into realistic medium-urban LiDAR scans to simulate occlusion scenarios and investigate the effect of the point cloud occlusion on the map-based NDT scan matching method performance. Under the occluded situations, the induced positioning error is found to be positively correlated to the change of heading angle. Significant 3-D localization errors peaks, up to 42.41 cm, are identified repeatedly at circumstances while the LiDAR encounters a substantial change of yaw angle, and these error peaks amplify as the occlusion rate increases.