The Research Of RGB-D SLAM Algorithm Based On Geometric Features Loop Closing Detection

In recent years,intelligent robots are widely studied in the world.Mobile robot technology,which is considered to have an unlimited development prospect due to its strong application,is one of the most important sub-fields of intelligent robots.Simultaneous Localization and Mapping(SLAM)is the key technology to achieve the autonomy of mobile robots.At present,the camera type sensor,Kinect as the representative,is widely used in visual SLAM research due to its low price and strong scene reconstruction.The main problems of the current visual SLAM technology are:(1)the issues of real-time caused by the long image processing time;(2)poor accuracy of visual caused by noise.In this thesis,we have optimized the key processing time and precision of the improved SLAM system based on the existing RGB-D SLAM algorithm.The main work of this paper mainly has the following points:(1)The SiftGPU algorithm accelerated by GPU is used to replace the Sift algorithm in the original algorithm.And the thesis also compared the SiftGPU and ORB extraction algorithm in efficiency.(2)In the Visual Odometry of SLAM,the RANSAC algorithm with depth information is used to reduce the false matching in feature matching to obtain a more accurate interior point matching pair.The ePnP algorithm is used to obtain the pose movement result of the robot in 3-dimensional space.(3)In the SLAM back end framework,loop detection is usually used to reduce the cumulative error of robot trajectory.Aming at the complexity of the trajectory,the geometric features of mobile robot's trajectory,which is average curvature K of the current trajectory is introduced based on the traditional circle detection algorithm.Calculate the K of each movement,and the different parameters of the mixed loop detection are accomplished according to the range of K value.It improves the robustness of the local motion trajectory by matching more near-loop loops when the curvature is large.(4)The continuous error of the robot pose calculation in the existing RGB-D SLAM system,which results in the visual odometry with the lost situation not being restored,has been considered.The recovery system proposed in the paper can achieve the automatic recovery when the visual mileage meter is lost.(5)For the cloud point map for the original algorithm which is susceptible to noise,unable to connect directly,and the larger storage when the map is established,Octomap is used as the navigation map,which converts the 3D Octomap map to 2D navigation map.Based on the RGB-D SLAM system and the famous TUM visual data set,the SLAM front-end processing speed is raised to 0.051 s.with the help of improved loop detection algorithm proposed in this thesis,SLAM back-end accuracy increases by 30.8% maximally,and by 18.4% on average.our fast RGB-D SLAM algorithm based on intelligent loop closing detection made contribution to the construction of real-time high-precision SLAM system.