Semantic mapping with image segmentation using Conditional Random Fields

Most maps used in navigation by mobile robots represent only spatial information. By the other hand, semantic information, which could be thought of as the classification of spatial primitives in different classes, provides structure to spatial information, hence reducing any necessary computation over the final map. This article proposes a semantic mapping process that represents an association between obstacles in a grid-based map and the correspondent regions of interest (ROI) in images from a vision system. The implementation consists of clustering laser measurement points related to a single obstacle and projecting them in images to be segmented using a Conditional Random Field (CRF) model to obtain visual descriptions of the detected obstacles. Results with real data are provided.