Indoor Environment Semantic Topological Mapping Based on Deep Learning

Abstract

With the popularity of mobile service robots in the home environment, people have put forward higher requirements on the robot's environmental perception. In order to improve the semantic ability of service robot perception of indoor environment, this paper presents a real-time semantic mapping system based on scene learning. Deep convolutional neural network is used to identify indoor typical scenes without training in a specific environment. Then, combining with vision and laser range data to estimate the semantic region, a semantic map layer is building based on a metric map. Furthermore, in order to express the indoor environment concretely, each semantic area in the room is subdivided into several functional areas as semantic topological nodes into the semantic map, thus constructing a semantic topological map. Finally, the feasibility of the semantic topological mapping algorithm is verified through experiments in real indoor environments.