Building Fast Dynamic 3D Maps for Trajectory Planning of Autonomous Ground Vehicles Using Non-Repetitive Scanning LiDAR Sensor

Building fast and reliable maps of the environment is a fundamental task for autonomous navigation. However, this process offers several challenges such as accurate registration of 3D point clouds. Recently, non-repetitive scanning LiDAR sensors have emerged as a promising alternative for 3D data acquisition, leveraging some of these challenges. In this paper, we present a 3D point cloud registration method that exploits the unique scanning pattern of such a sensor to register successive 3D scans. Errors accumulated over larger distances due to drift as well as registration errors due to the presence of dynamic objects in the scenes were improved by re-enforcing the registration method by first segmenting and classifying static and dynamic objects by analyzing the deformation in the unique, non-repetitive, Spirograph-type, scanning pattern of the sensor and then incorporating a fast NDT (Normal Distribution Transform) based registration method as well as loop closure detection for fine alignment. The novel method is then extended to build and update fast dynamic maps of the environment for trajectory planning for autonomous navigation. The proposed method is evaluated on three real and different datasets and compared with other state-of-the-art methods. The results not only demonstrates the suitability of these types of sensors for such applications but also show that the proposed method is comparable with other methods in terms of accuracy and surpasses them in performance in terms of processing time, making it suitable for real time applications.