Research Of SLAM For Cloud Robotics Based On RGB-D

Simultaneous Localization and Mapping(SLAM)problem is considered as the key to autonomous navigation for mobile robots.With its low cost and abundant scene information,SLAM based on visual sensor has drawn the researchers' attention gradually.Visual SLAM is a typical compute-intensive task,and the traditional solution depends on the computing resources of the robot itself,which has the disadvantages of poor execution efficiency and high requirements on robotic hardware.The concept of cloud robotics provides a new way to solve the above problems.With the help of powerful computing and storage capabilities of cloud platform,this thesis researches of SLAM system for cloud robotics based on RGB-D.In order to reduce the data transmission between local and cloud,it is necessary to extract the key frames from the RGB-D image sequence captured by the Kinect.Because the processor used in this thesis is a low-cost development board,its processing capacity is limited,and the conventional feature matching algorithms cannot meet the real-time requirements.In this thesis,the direct method is used to track the FAST feature points,and the process of calculating and matching the feature descriptors is omitted.After estimating the pose information of the mobile robot,this thesis sums the vector norm of rotation and translation as the criterion for selecting key frames.Experimental results show that the proposed method can improve the real-time performance and accuracy of key frame selection.In network communication,in order to ensure the continuity and reliability of data transmission,this thesis designs a Socket-based local-cloud data transmission scheme.On the robot side,once a key frame is extracted,it is transferred to the cloud database via Socket.The tests on public cloud and private cloud show that the data transmission scheme designed in this thesis can effectively reduce the occupation of network bandwidth and ensure the real-time transmission of key frames.In this thesis,we launch an instance in the cloud according to the actual needs,and eliminate the cumulative error by ORB-SLAM algorithm to optimize the key frame pose.The multi-threaded structure adopted by ORB-SLAM can't run on embedded processors in real time,and its sparse feature map can't satisfy navigation and interaction functions.Therefore,based on the ORB-SLAM in the cloud,this thesis builds a dense point cloud map and an Octomap that can represent the space occupancy information,which lays a foundation for robot autonomous navigation.The experimental results show that the proposed system can effectively unload complex computing in SLAM to the cloud,which can reduce the hardware requirements of mobile robots,and make it possible to apply advanced algorithms in low-cost systems.