Implementation Of SLAM And Demonstration Verification System Based On Multi-sensor Fusion

In recent years,with the increase in mining demand,secondary mining of mines has gradually become a research boom.Due to the limitations of early equipment and technology,many mines still have rich minerals,but the waste mines are old and the internal environment is unknown.What's more,it may be accompanied by toxic gases and hypoxia.Therefore,waste mines should not be directly mined.For this special environment,Simultaneous Localization and Mapping(SLAM)can realize carrier positioning and mine map construction,which provides guarantee for the secondary mining of waste mines.At present,SLAM can be divided into laser SLAM and visual SLAM.The former has a good local positioning effect but no color information,and the latter has a good global positioning effect but is easily affected by light.Comprehensive analysis shows that there are deficiencies in the SLAM solutions that rely on a single sensor.Therefore,in order to adapt to the changing light,lack of texture,and occluded roadway environment,this thesis conducts a multi-sensor fusion SLAM roadway trolley research.First,the SLAM framework of multi-sensor fusion is built,which is based on the classic SLAM framework.The multi-sensor observation model of RGBD camera,lidar and inertial measurement unit(IMU)is established,and the SLAM problem is described mathematically.Then,visual SLAM front-end,laser SLAM front-end,back-end optimization and loop closure detection are studied.In the visual SLAM front-end,SIFT,SURF and ORB features are analyzed,combined with the experimental results,SURF features with moderate performance are selected as the feature extraction algorithm in this thesis.For feature matching,I proposed an adaptive NNDR matching algorithm based on Manhattan distance,which can further improve the matching accuracy and reducing the amount of calculation.At the front end of laser SLAM,ICP is used for point cloud matching,and accurate pose estimation of lidar is obtained.In terms of loop closure detection,a detection method based on visual features is adopted to reduce the cumulative error of SLAM.BA is used in local optimization,and pose graph is used in global optimization,which further reduces the estimation error.In order to further improve the positioning accuracy,a multi-sensor fusion pose estimation scheme is proposed in this thesis.The IMU and the odometer pose are fused with the extended Kalman filter algorithm,and used as the initial pose of the camera and lidar to achieve a preliminary improvement in the positioning accuracy of the system.Then,the visual odometery,lidar odometery and initial fusion pose are fused twice,and square root cubature information filter is proposed for nonlinear high-dimensional fusion.Simulation verification was carried out in conjunction with Gazebo.Simulation shows that the positioning error of the proposed multi-sensor fusion pose estimation method is smaller than that of pure vision and pure lidar methods.Finally,a multi-sensor fusion SLAM demonstration and verification system based on RGB-D camera,lidar and IMU was built and the program was verified.First,a multi-sensor joint calibration was carried out,and the static conversion between each coordinate system was obtained.Then,the SLAM scheme of multi-sensor fusion was verified in three different environments.Experiments show that the multi-sensor fusion SLAM scheme in this thesis is superior to the pure visual SLAM scheme and the pure lidar SLAM scheme in terms of positioning accuracy and mapping accuracy,and it can more accurately locate and reflect the real environment.