Semantic Synchronously Localization And Mapping Based On Deep Learning

To meet the demand of 3D semantic mapping for indoor robots in the future,this paper considers how to implement them from the perspectives of autonomy,scalability and usability:1)Robots should be able to autonomously perform Simultaneous Localization and Mapping(SLAM)using RGB-D image sequences;2)Robots should be able to detect,classify and identify object instances from the environment,and consequently be able to construct a global consistently semantic map.3)Robots should be able to persist,organize and dig out indoor environmental information,and thus be able to understand concepts in given task description from human and realize the harmonious co-existence with human beings.This paper explores these three above-mentioned sub-problems and explores a new way for robots to autonomously construct indoor semantic maps using RGBD sensors.Specifically,the main work of this paper includes into the following three aspects.First,by means of theoretical analysis combined with the mathematical model derivation,the visual SLAM technology using RGBD images and its implementation are introduced in detail.This research employs inverse depth in optimization for data fusion.Then the open dataset is used for algorithm evaluation,and the real environment of the laboratory is used for experimentation.Second,using the Mask-Region with CNN feature(Mask-RCNN)and graph cut technology,a single image is labeled and a semantically labeled map is built.Furthermore,an object-oriented semantic map is constructed based on the idea that coupling SLAM technology and object recognition technology.Finally,the ontology model is established,and the instance attributes with relative reasoning rules are defined.Using the semantic map,ontology-oriented semantic knowledge base is consequently formed,helping co-existence between human beings and robots.The main contributions and innovations of this paper include the following three aspects.First,this paper,starting with visual SLAM technology and combined with deep learning technology,constructs labeled three-dimensional voxel model of map.Second,this paper explores the concept,that SLAM is capable to help semantic and SLAM improves semantics in turn.Third,after constructing object-oriented semantic map.this paper explores the way to restore,organize and dig out semantic information.This research has certain reference value and significance for improving the intelligence level of indoor service robots.