Research And Implementation Of Mapping Technology Based On The Fusion Of Laser And Vision

Simultaneous localization and mapping(SLAM)technology is a key technology in the field of unmanned vehicle environment perception,among them,3-D map construction is the focus of topic.At present,the 3-D maps based on large area of outdoor construction are mostly sparse maps or semi-dense maps,which are still lacking in complete reconstruction of the environment.Aiming at the problem that it is difficult to construct outdoor dense maps,a set of outdoor mapping system that integrates solid-state lidar depth information and monocular camera visual information is researched and designed in this thesis.In this thesis,a mapping scheme based on sensor fusion is proposed,and an automatic calibration module is designed to achieve accurate calibration of lidar and camera data.The data processing module is designed to realize the acquisition and fusion of sensor data.The positioning mapping module is designed to realize 3-D dense map construction and motion estimation of unmanned vehicle.The hardware and software platforms are assembled,and the effectiveness of the algorithm and the advantages of the system in outdoor drawing construction are verified by testing and experiments in real scenes.In this thesis,an automatic calibration scheme of lidar and camera is developed to solve the problem that the external parameters are not reliable and cannot be corrected in time due to the scene change and sensor offset.The off-line calibration and online calibration were combined to extract and match Canny edges,and improved calibration objective function was constructed to optimize external parameters,which was suitable for outdoor scenes without human intervention.Aiming at the problem that the fusion data of dense map is not compact,RGBD key frame is constructed.The radar point cloud without motion distortion is projected into a depth image by using the external parameter model,and then the pixel points of the camera image are cyclically traversed and the exact depth values of the corresponding points are estimated to achieve pixel-level fusion,and the map construction is more consistent.To solve the problem that the system runs outdoors for a long time,the accumulated error is large,and the map is optimized.Based on the DBo W2 word bag model,the similarity between current data and historical data is compared to correct the global map,eliminate drift error and improve the robustness of the system.The measured results demonstrated that the system hardware synchronization accuracy can reach microsecond level,the absolute error of sensor registration is within a few pixels,the positioning accuracy can reach decimeter level,and the mapping results are clear and complete,with reliability and real-time performance.