Probabilistic Semantic Occupancy Grid Mapping Considering the Uncertainty of Semantic Segmentation with IPM

An occupancy grid map considering only geometric information is often used for autonomous mobile robots. There are various areas we do not want autonomous robots to enter outdoors, such as grass areas. These areas are not reflected in the occupancy grid map because geometric information is not sufficient to distinguish these areas. This work attempts to add semantic information about the ground surface to a prior occupancy grid map for recognizing traversable regions. We create a semantically segmented bird's eye view (BEV) using semantic segmentation and inverse perspective mapping (IPM) and then apply a one-sided truncated Gaussian filter and binary Bayes filter to deal with the uncertainty of semantic segmentation and IPM. We tested our method on an approximately 1-km route at the University of Tsukuba and found that the recognition accuracy is highest if we apply these two filters together.