Research And Implementation Of Indoor Map Construction Algorithm Based On Semantic Segmentation

In the field of mobile robots,autonomous driving and high-precision mapping,an important requirement is the precise definition of the surrounding environment.SLAM(Simultaneous Localization And Mapping)has always been a key technology for sensing And locating the surrounding environment.With the increase of the complexity of robot application scenarios,indoor robots in particular need to understand the semantic information of the surrounding environment to achieve more intelligent services,and it is increasingly necessary to build a semantic map with accurate semantic information about the surrounding obstacles.At present,the main SLAM algorithms focus on the relative relationship of obstacles,and the semantic recognition of obstacles usually depends on the later annotation.Therefore,it is necessary to research and implement realtime semantic map construction algorithm for mobile robots,especially indoor robots.In this thesis,based on ESANet,an indoor real-time semantic map construction algorithm based on RGB-D semantic segmentation is proposed by studying mainstream visual SLAM algorithms and semantic segmentation networks.Meanwhile,in odometer tracking,a reprojection error correction algorithm based on semantic constraints is proposed.Based on the pixel semantic information and original depth images provided by semantic network,a semantic point cloud generation algorithm based on pixel semantic data and point cloud center of gravity is proposed.According to the results of the algorithm,a semantic map building system for indoor mobile robots is designed and implemented according to the operating environment characteristics of indoor mobile robots.After the actual deployment test,the indoor robot semantic map construction system based on the results of the algorithm research,in the data set and the actual environment,completed the task of constructing accurate semantic map.It improves greatly in odometer error correction and point cloud semantic recognition in fuzzy region.In the actual robot environment deployment,the system also ensures the real-time operation in the indoor environment.The test shows that the system has the ability to construct accurate semantic map of mobile robot’s surrounding environment information in the indoor environment,and has great significance for the existing indoor robot products such as sweeping robot and food delivery robot to further improve their intelligence degree.