Research on vector map SLAM for autonomous driving via camera and 4D radar

Abstract

Today, many companies and institutions are conducting research on autonomous driving and are about to be commercialized. In most cases, GPS is used to estimate the global localization for autonomous driving, but when GPS is used alone, there are problems that do not operate in environments such as multipath phenomena, signal rejection by high buildings, and tunnels. Therefore, to complement this, a vector map containing lane information for ego vehicle localization is essential. When creating a map, camera, LiDAR, and radar, which are representative visual sensors, are used. While most mapping algorithms are dependent on LiDAR, but its high cost poses a barrier to commercialization. Moreover, the creation of vector maps necessitates a considerable amount of time and effort. Consequently, this research aims to develop a real-time vector map SLAM with lane information utilizing cost-effective camera sensors and the 4D radar sensor which is currently under the spotlight. In the first research chapter, an 4D radar dataset in urban areas with various odometry sensors and calibration method will be proposed. In the second research chapter, monocular visual odometry using the Dynamic Objects Removal Mask (DORM) is proposed to enhance robustness in dynamic environments and improve performance. In the third research chapter, loop detection and pose graph optimization are conducted based on the 4D radar data and the generated vector map, aiming to minimizing trajectory drift errors.