Particle-Based Dynamic Semantic Occupancy Mapping Using Bayesian Generalized Kernel Inference

Abstract

A representative and accurate environment model is essential for the safe navigation and operation of intelligent transportation systems, such as autonomous vehicles and mobile robots. This paper presents a semantic occupancy grid mapping approach that uses a particle-based map representation to approximate continuous dynamic environments. The proposed approach recursively updates occupancy, velocity and semantic class estimates using the Bayesian Generalized Kernel Inference (BGKI) framework to maintain a local occupancy map in real time. The novelty of this approach lies in its combination of the continuous static semantic mapping capabilities of BGKI with the recursive dynamic state estimation of Dynamic Occupancy Grid Maps (DOGMs) in the 3D domain. We demonstrate that the approach maintains the semantic mapping capabilities of BGKI while providing more accurate velocity estimates than previous particle-based three dimensional DOGMs on real and simulated automotive datasets, including Semantic KITTI. We show that our approach outperforms the current state of the art on both semantic mapping and velocity estimation.