Research On Simultaneous Localization And Semantic Mapping Of Mobile Robots In Complex Scenes

Simultaneous Localization and Mapping(SLAM)technology is a key technology for mobile robots to locate and map in unknown environments.With the development of computer vision technology,vision-based SLAM technology has attracted widespread attention due to the advantages of low-cost visual sensors and rich information.However,as the application scenarios of mobile robots become more and more complex and the tasks become more and more intelligent,the traditional visual SLAM technology can no longer fully meet the current needs.In order to enable mobile robots to obtain precise positioning and richer semantic information in complex environments,this paper constructs a simultaneous positioning and semantic mapping system for mobile robots in complex scenes.The main contents are as follows:(1)Aiming at the complex work scenes and intelligent tasks of mobile robots today,this paper uses multi-concurrency technology to add dynamic feature elimination and semantic map construction on the basis of ORB-SLAM3,and proposes a complex scene.Simultaneous localization and semantic mapping system.The system can eliminate the impact of dynamic targets in the environment on the positioning and mapping of the visual SLAM system,and build an environmental semantic map containing instance-level semantic information.(2)In order to eliminate the impact of dynamic objects existing in complex scenes on the system,this paper designs a dynamic feature culling method based on the tight coupling of instance segmentation and multi-view geometric constraints.Segment the scene with the SOLOv2 instance segmentation algorithm that takes into account the segmentation speed and accuracy,and obtain the prior information of the target motion;use the multi-view geometric constraints to detect the real dynamic features;finally,use the tight coupling method to make full use of the motion prior of the instance segmentation The complementarity of motion detection information and multi-view geometric constraints enables accurate removal of dynamic features in the scene.(3)In order to satisfy mobile robots’ deeper perception and understanding of complex work scenes,this paper designs an environment map construction method incorporating instance level semantic information.Firstly,3D point cloud images are generated through point cloud generation,stitching,and filtering;Then,the instance segmentation results are used to extract the 3D semantic point cloud of the environment target,generate the corresponding 3D semantic tags,and construct an instance level semantic tag library;Finally,a global environment instance level semantic point cloud map is constructed through the fusion and optimization of point clouds.In order to facilitate the storage and update of the map,a dense octree semantic map is further constructed,thereby completing the construction of the environment semantic map.