Research On Key Technologies Of Visual SLAM Under Incomplete Informatio

The map constructed by the traditional visual SLAM system can only enable the robot to perform simple tasks,such as path planning and navigation,but cannot achieve a deeper understanding of the objects contained in the constructed indoor environment map,so that the robot cannot perform more intelligent and complex tasks.Meanwhile,when the robot is in a weak visual environment,the robustness of the system will also be reduced,and even lead to the failure of mapping.The construction of semantic map can help robot achieve a deeper understanding of their indoor environment.Therefore,this paper enables to solve the problem of using RGB-D cameras to construct semantic 3D scene map in weak visual indoor environments and designs a system to construct the 3D semantic map which contains rich semantic information of indoor objects.The research work of this paper mainly includes the following contents:A light-weight semantic segmentation algorithm based on Light-Weight Refine Net.In order to solve the problem that the SLAM system has poor real-time understanding of scene semantics in practical application scenes,firstly,a light-weight semantic segmentation model Light-Weight Refine Net is used to segment images to improve the segmentation speed of scene images.Then,the Light-Weight Refine Net used in the semantic segmentation thread is experimentally analyzed,the network parameters are optimized,and a database of semantic label information is constructed at the same time,and finally a low-cost and high-precision semantic segmentation model is obtained.The input image is processed for real-time segmentation.A visual semantic 3D map construction method based on ORB-SLAM3.In order to solve the problem of the failure of map construction due to the loss of inter-frame tracking in the weak visual environment,firstly,the characteristics of map reconstruction of ORBSLAM3 are analyzed,and used as the underlying architecture of the visual semantic 3D scene mapping framework to provide steadily estimate of the camera pose to solve the problem of loss of inter-frame tracking information during semantic mapping.Then,the PCL point cloud library is used to obtain point cloud data containing structured information and the Light-Weight Refine Net network is used to extract the semantic information of scene objects.Secondly,the information fusion mechanism and method of the visual semantic 3D scene mapping system are designed and optimized.Finally,the Octo Map 3D scene reconstruction module is used to realize the 3D map representation containing semantic information,and finally a visual semantic 3D scene map construction method based on ORB-SLAM is formed.Visual semantic 3D mapping system integration and testing.Firstly,the designed visual semantic 3D scene mapping system is integrated on the platform of social robot of Rabbot then,the system was tested and analyzed in the public indoor scene dataset TUM and the real laboratory scene.