Research On SLAM Algorithm Fused With Stereo Visual Inertial Odometry

Simultaneous Localization and Mapping(SLAM)technology is of great significance for mobile robots to realize the perception and interaction of the environment.In complex environments such as too fast and texture-free areas,it is difficult for mobile robots to achieve high-precision positioning.Based on the fusion of binocular vision and inertial navigation information,a SLAM framework for stereo visual-inertial navigation based on the direct method is proposed.Firstly,in order to solve the problem that the extraction of sparse feature points by the feature point method affects the positioning accuracy,an inertial measurement unit(IMU)assisted feature tracking method based on the direct method is proposed.Using image pixel gradient changes,a large number of area blocks are down-sampled to extract feature points.In order to better use the information of the point,the selection method of the point is expanded to construct an eight-dimensional residual and share the depth information of the point.In the feature tracking process,according to the advantages of IMU short-term estimation,the measurement information of the IMU is used to rotate the point,and the outliers will be eliminated and filtered in the next frame projection position to reduce false matches.In addition,in order to improve the accuracy of key frame selection,in addition to adding traditional perspective changes when making key frame decisions,occlusion and exposure time constraints are also added to avoid redundant key frames consuming system resources.The experimental results show that the outliers can be well detected after adding IMU measurement information,which improves the positioning accuracy of the system.Secondly,in view of the problem that the feature points generated in the pure vision algorithm due to too fast motion or the lack of texture areas are unable to perform good feature tracking and pose estimation,a joint optimization method fused stereo odemotry(VO)with IMU are proposed(stereo_VI_DSO).A target energy function combined with a photometric error function and an IMU preintegration residual is established to reduce the complex of the computation,the form of sliding window is used to perform the real-time system,and the information of marginalization as a priori information of the target energy function.Iterative optimization is performed on the external parameters,posture,and speed of the camera and IMU using the pose optimization method.The proposed algorithm is tested in indoor dataset EuRoC,outdoor dataset KITTI and indoor actual scenes.The experimental results show that the algorithm in this paper can be positioned and framed in real time in a complex environment,and the positioning accuracy is higher than that of monocular inertial navigation DSO(Mono Visual-Inertial Direct Sparse Odometry),OKVIS(Keyframe-based Visual-Inertial System)and feature point method inertial navigation scheme VIORB-SLAM,and the proposed method has higher reconstruction density and scene recovery ability.Finally,aiming at the problem of repeated calculations in closed-loop scenes,due to closed-loop scenes and height relocation in stereo_VI_DSO,a closed-loop detection method with semantic topology map is proposed.The Mask R-CNN neural network is used to segment the input image to obtain semantic labels,and map to topological nodes and point cloud maps to obtain semantic topological maps.The extracted Shi-Tomasi corner points are combined with the original feature points,and the ORB descriptors of the corner points are calculated to form reusable feature points,and the loop detection system is constructed using semantic nodes and bag-of-words models(Bag-of-Words).The semantic topology map is combined.Experimental results show that the proposed algorithm can construct a clear semantic topology map and the positioning accuracy is higher than stereo_VI_DSO and closed-loop detection DSO(LDSO).