Deep occupancy maps: a continuous mapping technique for dynamic environments

Learning a model of an environment that is correctly able to distinguish occupied and unoccupied areas is important for maneuvering robots in unstructured environments. A common technique to tackle such problems is to train a classifier with hand-crafted features that encode occupancy information. However, finding good features quickly becomes computationally prohibitive and impractical for complex and large environments. In this paper, we propose a “fully” convolutional neural network which can build global continuous occupancy maps by learning from raw local unorganized point cloud data, conveniently captured using a range sensor such as LIDAR. We propose a novel way of, 1) transforming data into a grid representation, and 2) perform convolution on robot’s position rather than on occupancy levels. With this formulation, the map can produce both static and long-term maps in large environments without altering the model. Since the model is a function of locations, it is possible to query the occupancy probability at any position in the environment. Experiments indicate both computationally-efficient and accurate results over other continuous occupancy mapping techniques that require manual feature extraction.