Simultaneous localization and mapping based on the local volumetric hybrid map

Abstract

Simultaneous localization and mapping (SLAM) plays a significant role in autonomous vehicles when a global navigation satellite system (GNSS) is not available. Environment models and underlying estimation techniques are key factors of this algorithm. In this paper, we present a hybrid map-based SLAM approach using Rao-Blackwellized particle filters (RBPFs). We represent the environment with the hybrid map which consists of feature and grid maps. The joint posterior between the vehicle positions and both maps are maintained using RBPFs. This approach allows a vehicle to update its states in a more robust and efficient way. We derived a novel sampling formula by combining a feature measurement likelihood to the traditional grid-based SLAM framework and can decrease the uncertainty of the predicted vehicle position significantly. Moreover, we represent the grid maps with 3D models because 2D models could be insufficient and less reliable to achieve tasks such as navigation and obstacle avoidance in complex 3D environment. We are also able to show that the 3D grid measurement likelihood has a lower variance and with that we can improve the overall performance of the algorithm.