| Simultaneous Location And Mapping also known as SLAM,is the core technology in autonomous navigation of mobile robots,augmented reality,virtual reality and other popular fields.The technical problem solved by SLAM is how to calculate pose of the object and construct the map information of surrounding environment in parallelly.The camera has been extensively researched and adopted in academia and industry because of rich information.In the common monocular vision SLAM,the matching accuracy of feature points is generally not high,the method of selecting keyframe is improper,and the error of the front-end and back-end optimization algorithm is large.The overall algorithm cannot meet the requirements of real-time.This paper will mainly study the methods of feature points and keyframe selection during SLAM operation,improve the front-end and back-end optimization algorithm and solve the real-time problem of SLAM.Studying these key steps has important theoretical guiding significance for improving the localization accuracy of SLAM,and is an important guarantee for the accuracy and real-time of SLAM.The main research contents of this paper are as follows:(1)In the front-end scheme design,the extraction steps,characteristics and matching principle of the popular feature points are studied.The pose of camera and space point estimation method and the monocular camera initialization method are discussed.A keyframe selection strategy is adopted to ensure richness and effectiveness of image frame content and can also help reduce the amount of calculation and avoid failure while tracking in the SLAM process.The loop detection method based on the ORB bag of words model quickly helps determine the position of the loop and avoid wrong estimation caused by cumulative error.The error estimate has greatly improved the localization accuracy of SLAM.Finally,the overall design scheme of the monocular vision SLAM front end is given.(2)In the back-end optimization step,two different optimization ideas are first introduced,and the overall optimization method is determined.The two main methods are compared.Then,the Newton method is used in the g2 o library to compute the local,global and loop optimization to optimize the pose and space points.(3)The effects of feature point extraction and matching were studied from the aspects of illumination intensity,image rotation,image blur,scale change and weak texture area.The experimental results show that the ORB feature points have excellent matching performance.The keyframe extraction method of time domain,spatial domain and image content method is studied experimentally,it shows that the selection method based on image content has excellent performance.Using one-dimensional quadratic equation plus noise to simulate the interference of SLAM during the operation,which shows that the Levin-Marquart method has more accurate optimization results.(4)The experimental research on the localization accuracy is carried out.The experimental results show that the maximum root mean square error of the absolute trajectory error and the relative pose error of the camera is 0.20336 m,and the maximum root mean square of the rotational error of the relative pose error is 3.4054 deg.It is smaller than the corresponding total translation and total rotation.The results show that the SLAM algorithm can basically recover the actual trajectory of the camera.(5)Finally,a SLAM method based on wireless network is studied.After experimental verification,the function of localization and mapping on embedded devices is realized.The calculation speed is improved,the real-time performance is enhanced,and solved the problem of insufficient computing power of mobile devices.Figure [58] table [13] reference [51]... |
PDF Full Text Request