Sparse Visual Semantic Point Cloud-Based Relocalization Method for Intelligent Vehicles

Localization of intelligent vehicles often depends on a combination of RTK-GNSS/IMU (Real-Time Kinematic-Global Navigation Satellite System/Inertial Measurement Unit) and high-definition maps. However, production vehicles can’t burden the high cost of RTK-GNSS/IMU localization systems and high-definition maps. Scenes like Viaducts, buildings, and tunnels where the signals from satellites and ground stations are absent may have a negative influence on the localization performance of intelligent vehicles. In this paper, a low-cost sparse visual semantic point cloud-based relocalization method that leverages deep learning techniques is proposed to tackle the above issues. Our method extracts highly representative descriptors from the visual point clouds based on the deep learning network. The framework of our proposed method comprises a joint localization mechanism combining global and local matching achieving relocalization without relying on RTK-GNSS/IMU and highdefinition maps. And a temporal successive matching localization mechanism is proposed for initialization. Achieving a low-cost relocalization solution for intelligent vehicles by using a lightweight and sparse visual semantic point cloud. The visual point clouds used for experiments are collected in two different urban areas in Chongqing, comprising 33 and 20 rounds respectively. We verified that our method has superior accuracy and robustness after comparing and analyzing the proposed method with the classic deep learning-based relocalization method on the self-collected dataset.