Arbitrary vehicle localization in indoor parking lots using infrastructure-based RGB-D camera

Abstract Automated valet parking is an important application scenario of intelligent driving technology. The core challenge of this application is to achieve accurate and robust vehicle localization and perception in the parking lot. In recent years, vehicle-road collaboration technology has developed rapidly. Compared to the agent-based approach, the infrastructure-based intelligent solution equipped with uncentralized sensing devices and computing platforms has the advantages of low overall cost and unobstructed view, which makes it widely concerned in the field of intelligent driving in limited scenarios represented by indoor parking lots. In this paper, we use the method of arranging RGB-D camera arrays on the infrastructure to build an indoor parking lot localization system and then focus on the two difficult problems of vehicle online 3D modeling and vehicle pose estimation. Firstly, aiming at the cumbersome offline generation of vehicle models, we propose an online automatic generation method of vehicle point cloud models based on LiDAR with non-repetitive scanning. Secondly, aiming at the problem that the 3D object detection method is difficult to achieve accurate vehicle pose estimation, we propose a local template registration method for conditions where RGB-D cameras cannot measure low-reflectivity objects. Finally, a complete hardware system is built to verify the feasibility of our solution in the real-world environment. The experiments in an indoor parking lot demonstrate the universality and accuracy of the proposed vehicle localization system.