Real-Time Driving Scene Understanding via Efficient 3-D LIDAR Processing

Wonje Jang, Minseong Park, Euntai Kim

Abstract: The 3-D light detection and ranging (3-D LIDAR) sensors are widely used in autonomous vehicles; however, their drawback is the significant computation processing requirement. Although parallel processing using high-end graphics processing units (GPUs) can solve the aforementioned problem, the cost of high-end GPUs prevents 3-D LIDAR from being used in commercial vehicles. In this study, a new fast and efficient perception method using 3-D LIDAR is developed. The aim is to develop a real-time perception method that can run on both a low-end GPU and a central processing unit (CPU). The proposed method consists of three parts: preprocessing, point cloud classification network (PCCN), and filtering. To process a point cloud using minimal computation and achieve good performance, handcrafted features and learning-based features are used in PCCN—the former capture the global information of the given point cloud, whereas the latter encode the local information of the points in the cloud. Finally, the proposed LIDAR perception method is applied to the SemanticKITTI dataset, and it is validated by comparing with other competing methods in terms of accuracy and computation time on the CPU.