Research On Simultaneous Localization And Mapping Of Unmanned Vehicles In Complex Indoor Environment

Nowadays,SLAM(Simultaneous Localization and Mapping)technology has been applied to many fields such as military,education,medical care,etc.Among them,the combination of SLAM technology and unmanned vehicles is a research hotspot.Based on the visual slam system,this paper studied the simultaneous localization and mapping of unmanned vehicles in complex environment,and designed a complete visual slam scheme.The main work contents were as follows:Firstly,aiming at the problem that visual SLAM system was difficult to maintain real-time and stability in complex scenarios,an improvement method was proposed to improve the ORB(Oriented Fast and Rotated BRIEF)feature points that were too dense,improved the traditional SLAM framework.The front end used the division of pixels combined with the quad tree to complete feature extraction,the pixels were divided to improve the probability of some areas extracted to the feature points,and the quad tree method uniformly distributed the feature points for extraction.The combination of EPNP+ICP by RANSAC reduced the error of solving the camera motion,the traditional algorithm and the improved algorithm were compared experimentally under the TUM dataset to verify the improvement of the real-time and stability of the improved algorithm.Secondly,in view of the adverse effects of dynamic feature points on SLAM positioning and mapping in complex scenarios,this paper studied the removal of dynamic feature points,first used YOLOv3 to detect and correct potential dynamic objects,and then used motion consistency to divide static dynamic regions,and classified feature points in the area,and then based on optical flow method for motion detection and tracking,preliminary judgment of the motion state of objects.The re-evaluation and validation of the pole constraints were then combined to solve the problem of too many false rejects leading to loss of positioning.Finally,combined with the dynamic information,the dynamic feature points were eliminated to reduce the influence of the dynamic feature points on the pose estimation,improved the positioning accuracy and system stability,and used the dynamic data set to verify that the method in this paper was more feasible and stable.Then,the back-end,closed-loop detection and mapping of visual SLAM system were studied.The closed-loop detection scheme based on Bag-of-words model improved the efficiency and accuracy of closed-loop detection.Based on the G2O image optimization model,the pose of the camera was optimized,which reduced the computation and improved the realtime performance of the system.Finally,the map construction and map types were studied,and the experimental results were analyzed to verify the superiority of the back-end part.Finally,the experimental platform of the unmanned vehicle was studied,including the hardware system and software operating system of the unmanned vehicle,and the feasibility of the proposed method was verified on the experimental platform.The analysis of experimental results showed that the proposed method effectively improved the real-time performance and stability of the system.