| Mobile robots have achieved large-scale applications in industry,military,medical and other industries.With the increase of application scenarios and application requirements,the research on the intelligence of mobile robot is also gradually carried out,including positioning,mapping,path planning and navigation.The ability of mobile robot to analyze and understand the scene is also an important standard of its intelligence.The introduction of semantic map plays an important role in promoting the production of high-precision map,automatic annotation in crowdsourcing map and robot positioning based on semantic information.By studying the relevant literature on mobile robot positioning,mapping and target detection at home and abroad,this thesis combines laser slam and target detection methods to build a point cloud map with semantic information.The main research contents are as follows:(1)The calibration method between lidar and camera is studied,and the off-line calibration of solid-state lidar and camera based on rectangular calibration plate is completed.The scheme of on-line calibration of non cloud lidar is combined with the characteristics of solid-state lidar.The experimental results are analyzed to verify the accuracy of the online calibration scheme,and realize the real-time correction of external parameters of camera and solid-state lidar in outdoor scene.(2)The SLAM algorithm of solid-state lidar is studied.According to the scanning characteristics of solid-state lidar,such as non repetitive scanning and small field of view,the point cloud distortion is corrected based on the uniform motion model,and the feature extraction method of point cloud data is designed.In view of the large amount of point cloud map data and excessive redundant calculation in the process of mapping,the key frame extraction scheme of laser odometer is designed,the factor graph optimization model based on point cloud key frame attitude and IMU pre integration attitude is established,and the sliding window method is used to fuse the solid-state lidar and IMU data.Experiments show that the odometer based on factor graph optimization has better positioning accuracy at turns and lower overall cumulative error.(3)Aiming at the problem that the sampling frequency of solid-state lidar and binocular camera is different and the binocular camera has no external hardware trigger,the spatio-temporal software synchronization and data fusion scheme of both are designed,and the area to be detected is divided according to the different FOV of both,which reduces the redundant calculation in the process of data fusion and realizes the construction of color point cloud map.(4)Using the yolov5 target detection algorithm,the yolov5 s neural network is trained through the self-made data set to realize the target detection of three common objects in the campus scene.The key frame point cloud data of laser inertial odometer is fused with the target detection results of camera,and the point cloud with semantic information is semantically segmented to realize the construction of three-dimensional semantic map in outdoor scene.The research of this subject is helpful for the large-scale production of highprecision maps,the automatic annotation of large quantities of point cloud data,and the human-computer interaction of mobile robots in the field of automatic driving,and lays a certain foundation for the positioning of mobile robots based on semantic information. |
PDF Full Text Request