Understanding a changing environment is one of the essential elements in autonomous robots and systems. A depth sensor or a LiDAR provides the geometric information of the surroundings as a point cloud. As a fundamental technique for robotics applications such as motion planning and collision avoidance, occupancy mapping techniques from sensor data have been studied for decades. However, updating robust representation with real-time processing speed remains a challenging problem of occupancy mapping in a dynamic environment, since the point cloud data of a single scan contains partial geometry observations of the environment.
This dissertation studies the occupancy mapping techniques that update their dense occupancy representations in real-time, exploiting spatial correlation of point cloud data.
Specifically, we propose a method for real-time occupancy updates in a dynamic environment; 1) an acceleration algorithm exploiting the geometric update patterns of an occupancy map. Furthermore, we propose two approaches to update the dense occupancy representation of the environment given sparse sensor data; 2) regression method using correlation among occupancy observations, and 3) deep-learning network embedding prior knowledge about the spatial correlation of measuring sensor data.