Research And Implementation Of RGB-D SLAM Based On Key Frames

RGB-D SLAM uses RGB-D sensor as data input,and SLAM technology is used to simultaneous localization and mapping.Nowadays,SLAM is receiving more and more attention in indoor mobile robots,unmanned vehicles and drones.In this paper,the RGB-D SLAM algorithm is divided into front-end and back-end.Each module is researched and analyzed,and corresponding algorithm improvements are made,which makes the whole SLAM algorithm more robust and accurate.The research mainly includes the following aspects:1.At present,the depth error estimation of the RGB-D sensor is aimed at a specific depth camera,the disadvantages of these methods is that it is put forword for a specific sensor and cannot be applied to the other sensors.In order to solve this problem,this paper proposes a general method to estimate the root mean square(RMS)error of depth data provided by general three-dimensional sensors.The method is applicable to three-dimensional sensors based on structured light,time-of-flight,stereo vision and other technologies.Use a common checkerboard to detect corner points and get two point clouds,one is the real point cloud of the image corner,and the other is the estimated point cloud of the corner point given by the device.After registrating of the two point clouds,the RMS error is calculated.The RMS error is generalized as a function of the distance between the RGB-D sensor and the checkerboard.The accuracy and practicability of the proposed method are verified by comparing it with the existing advanced depth error estimation methods.2.An improved mismatch culling algorithm is proposed based on the feature extraction and matching algorithm.The experimental results show that the improved algorithm can retain more correct matching points while eliminating mismatch(for SIFT,SURF,ORB algorithms,respectively increased by 9.9%,16.7%,and 28%).while the usage time was also reduced(for SIFT,SURF,and ORB algorithms,they were reduced by 4.7%,25.1%,and 61%,respectively).3.Several links of the back end of RGB-D SLAM method are studied.Under the research of the existing key frame detection algorithm,an improved key frame algorithm is proposed,and the selected key frames are filtered and improved.The keyframe selection algorithm improves the robustness and precision of keyframe selection.At the same time,the closed-loop detection algorithm based on visual dictionary is used to solve the problem that the closed-loop detection efficiency of the system is low under long-term operation.Finally,use graph optimization to obtain a globally consistent camera pose and point cloud map.4.A visual SLAM framework based on RGB-D is constructed.The simulation experiment was carried out using the standard data set,and the absolute trajectory error ATE was introduced as the quantitative standard.The system was qualitatively and quantitatively analyzed.At the same time,in order to verify the use of the entire algorithm in the actual scene,the improved algorithm is applied to the actual scene.The above experiments demonstrate the robustness and accuracy of the improved SLAM system.