| Simultaeous Localization and Mapping(SLAM)technology is one of the most concerned research directions in recent years.This article focuses on indoor mobile robots,improves the existing classic visual SLAM framework and designs a SLAM algorithm suitable for indoor positioning,and verifies it through public datasets and the real environment collected by my camera,which significantly improves the speed and robustness of theindoor mobile robot SLAM algorithm.The main content of this article is as follows:1 Existing SLAM frameworks often consider most scenarios and do not take advantage of the environmental characteristics of indoor scenarios,resulting in low operating efficiency.Therefore,this article is first based on indoor horizontal ground constraints,various key points and descriptors are deeply studied,and designed related feature point experiments.The key points and descriptor combinations that are most suitable for the scenario in this article are obtained,which is five times faster than ORBSLAM2 on the front-end matching speed.Next,this paper improves its visual front end based on ORB-SLAM2,uses VO based on optical flow tracking,and preprocesses the image to improve the quality of feature points.Through experiments on the ICL-NUIM dataset,compared with ORB-SLAM2,the algorithm in this paper is nearly twice as fast,and the positioning accuracy is equivalent to ORB-SLAM2,and the algorithm is more robust in an environment where feature points are not abundant.2 Aiming at the defect that the monocular camera cannot estimate the absolute scale and the tracking loss in the indoor low-texture place,this paper is based on the improved visual front end,an algorithm based on sliding window fusion IMU is proposed,including the algorithm implementation of Tracking thread and Local Mapping thread,as well as the method of IMU sensor initialization;Study the error model and motion model of IMU,derive the IMU model under continuous time and discrete time,and compare the two integration methods,select the better median integration;use Kalibr to jointly calibrate the camera and IMU to obtain the external parameters between the IMU and the camera,which will lay the foundation for the subsequent system basis.Through experimental comparison in real scenes,the algorithm in this paper can avoid the problem of trajectory drift at indoor white walls;through experimental comparison under Euroc,the algorithm in this paper has higher positioning accuracy than pure visual SLAM methods and can solve the problem of uncertain scale.And it has the same robustness and positioning accuracy of the current advanced VIO system.3 Aiming at the impact of dynamic scenes in the room on positioning accuracy and relocation,an algorithm based on target detection and three-point frame difference method is proposed.Both algorithms are used to eliminate dynamic points.Based on target detection,Mark-R CNN is used.Based on three-point frame difference method,the geometric information of the image is used.Through the experimental compareson of TUM,the positioning accuracy of the algorithm in this paper is improved significantly in a high dynamic environment.The method based on target detection is improved by ten times,and the method based on three-point frame difference is improved by four times;under low dynamic environment,both are improved by 20%.Based on the three-point frame difference method is more suitable for the scene in this article due to its real time advantage. |
PDF Full Text Request