Revisit Data Association in Semantic SLAM Systems for Autonomous Parking

Revisit Data Association in Semantic SLAM Systems for Autonomous Parking

연구 분야: Artificial Intelligence

학회: International Conference on Multimedia Modeling

Semantic VI-SLAM (Visual-Inertial Simultaneous Localization And Mapping) systems are crucial to autonomous indoor parking as they perceive the environment by constructing a map with various semantic objects. A critical challenge in building such maps is the data association problem, which involves linking each currently observed semantic object with an existing landmark in the map. The primary issue arises from changes in the appearance of these objects as the vehicle moves, potentially leading to incorrect matches between observations and landmarks. In contrast, surround-view images taken from a top-down viewpoint offer a solution by capturing semantic objects on the ground that are both easily recognizable and consistent in appearance. Thus, in this paper, we propose a robust data association method for semantic VI-SLAM systems by incorporating ground semantic objects from surround-view images. By integrating both geometric and semantic measurements of surround-view objects in a tightly-coupled objective, our data association approach substantially improves in accuracy and robustness compared to traditional approaches. Moreover, metric information from surround-view images is used to refine the scale of the SLAM system, addressing the initial scale inaccuracies caused by the low precision of Inertial Measurement Unit (IMU) measurement. The efficacy and efficiency of our proposed data association method are validated in a typical real-world indoor parking environment.