The Research of 3D Point Cloud Data Clustering Based on MEMS Lidar for Autonomous Driving

Abstract

In the field of autonomous driving, the perception of the environment plays a crucial role, serving as a fundamental component. Accurate and precise environmental detection is vital in providing detailed information about obstacles for the control module of autonomous vehicles. MEMS LiDAR, as a prevalent sensor for acquiring obstacle positions, offers high accuracy in data acquisition by leveraging its dense point cloud information. However, a characteristic of MEMS LiDAR is the decrease in cloud density as the distance increases. Failure to consider this issue can lead to problems such as merging or splitting of obstacles during the clustering process. Furthermore, relying solely on a two-dimensional grid-based approach poses challenges when it comes to detecting overhanging obstacles. To overcome these challenges, we propose a method that tackles the problems of undistinguishable adjacent obstacles, splitting of distant obstacles, and the detection of overhanging structures. First, we apply ground segmentation techniques to remove ground-based points from the point cloud data. This step helps in isolating the obstacles of interest and improving the accuracy of subsequent analysis. Next, we create a three-dimensional grid map and determine the occupancy of each grid cell. To optimize the problem of distant obstacle splitting, we employ a dilation algorithm to expand the occupancy of the grid cells. Subsequently, we convert the three-dimensional grid into a two-dimensional representation and evaluate the occupancy of each cell in the resulting grid based on the height direction occupancy. Furthermore, we employ noise removal techniques to enhance the quality of the data. Finally, we utilize the DBSCAN algorithm, which incorporates an adaptive radius and eight-neighbor cells clustering algorithm, to perform obstacle clustering operations. Comparing our proposed method with the traditional DBSCAN algorithm, we observed that our method achieved a 7.6% increase in detection accuracy, while reducing calculation time by 16.2%.