Research On Monocular SLAM Technology Combining Direct Method And Indirect Method

Simultaneous Localization and Mapping(SLAM)is a technology that uses one or more sensors to locate itself in real time and construct an environmental map at the same time.It is widely used in mobile robots,autonomous driving,AR and other fields.In visual SLAM,SLAM can be divided into direct method and feature method according to whether the front end estimates whether the pose needs to extract features.In the direct method,it is not necessary to extract feature points,and a large number of keypoints information is directly used for pose estimation,and the precision is high and the speed is fast.However,the direct method requires a small distance between two frames.When the camera is shaken or dropped,the distance between the two frames is large,and the direct method is easy to track errors or failures.The key points in the direct method do not describe sub-information,which makes closed-loop detection difficult.Based on the direct method SLAM,this paper combines the direct method with the feature method to avoid the shortcomings of the direct method while using the feature information while ensuring the advantages of the direct method.The main contents of this paper are as follows:(1)Aiming at the key points in the direct method SLAM,the sub-information is not described,and long-distance matching is impossible,which leads to the difficulty of closed-loop detection.A fast closed-loop detection method is proposed.The method determines a closed-loop candidate frame by performing a similarity comparison on the historical frame window by using the threshold to determine a minimum similarity threshold in the current frame window.The experimental analysis of the proposed method and the current mainstream SLAM detection method proves that the detection method is faster and has better stability.(2)Aiming at the closed-loop correction problem in direct method SLAM,a global optimization correction method based on window common view is proposed.After detecting the closed loop,the method propagates the 3D point information in the window to the current frame and calculates a similar transformation between the current frame and the candidate frame,and then establishes a frame-to-frame error relationship according to the window common view information in the direct method,and The trajectory and map are corrected by global pose optimization.The experimental results show that the proposed method effectively reduces the cumulative error of the system.(3)Aiming at the problem of tracking error or failure when the distance between two frames in the direct method SLAM is large,a fusion pose estimation method is proposed.The method optimizes the 3D point information by inter-frame matching,and then performs feature matching and calculates the current frame pose by PnP method,and optimizes the pose by direct method.The experimental results show that the proposed method effectively improves the error of pose estimation when the distance between two frames is large.(4)Based on the several methods proposed in this paper,the SLAM system with fusion feature method and direct method is designed and implemented.The system mainly includes sensor input,data processing,fusion algorithm and several modules for positioning and mapping.At the same time,data collection is performed on a variety of real scenes,and it is run in the system,direct method DSO system and feature method ORB-SLAM system proposed in this paper.The experimental comparison and display of the running results show that the system has high accuracy of mapping and can effectively reduce system cumulative error.