Real-Time Occupancy Grid Mapping Using RMM on Large-scale and Unstructured Environments

Abstract

Occupancy mapping is crucial for distinguishing between known and unknown regions, which plays a significant role in the autonomous exploration of unmanned aerial vehicles (UAVs). However, the construction of high-quality maps is still a challenge. The challenge comes from the following factors. The vast amount of data captured by UAV exploration in large-scale environments brings computing and storage bottlenecks. Additionally, sensor noise and obstacle occlusions will affect the completeness of the map. To address these issues, this paper applies a lightweight mapping framework based on the Random Mapping Method (RMM) to the challenging task of real-time UAV exploration. This framework employs a linear parametric model, where RMM efficiently maps sensor data into a high-dimensional feature space, enabling rapid learning of occupancy states. We demonstrate that this approach is not only efficient in terms of computation and storage but is also particularly effective at inferring and completing unobserved map regions caused by sensor noise and obstacle occlusions. When the exploration is completed, the global occupancy grid map is stored and implicitly represented with limited parameters. Simulation and real-world experiments are conducted to verify their comprehensive performance compared to the typical methods and state-of-the-art methods.