Visual SLAM Localization Of Mobile Robot

In an unfamiliar scenario,the first task of a robot is to obtain location and surrounding information.The key technology is simultaneous positioning and mapping.Due to its application prospects in the field of robotics,the research on visual SLAM(Simultaneous Localization and Mapping)has attracted widespread attention.Based on the classic visual SLAM framework,this thesis studies the two modules of visual odometry and closed-loop detection.In the visual odometer module,in view of the poor robustness and scale drift problems of traditional methods,an optical flow visual odometer based on deep learning is proposed,which uses the excellent performance of convolutional neural network in image feature extraction to obtain more accurate images Matching relationship,thereby improving the accuracy of camera pose estimation,and optimizing the scale uncertainty of visual odometry.In the closed-loop detection module,most loop detections mainly use individual point features.However,in scenes with sparse point features,the performance of the closed-loop detection algorithm is poor.In addition to point features in the image,there are also line features from different sources of information.In the sparse texture scene,the line features in the image can appropriately supplement the point features.Therefore,in the closedloop detection module,the combination of point and line features is studied,and a closed-loop strategy combining point line features is proposed.The principal activities and contributions are as follows:First of all,in the traditional visual mileage calculation method,image matching takes a long time,image information utilization efficiency is low,and the environment is highly dependent.This article suggests an optical flow visual odometry derived from deep learning.Utilize the deep learning manner to estimate the optical flow of the image,thereby improving the accuracy of the two-frame image matching.At the same time,in order to further improve the accuracy of the algorithm’s pose estimation and optimize the scale uncertainty problem in the pose estimation,the visual mileage calculation The scale alignment module is introduced into the legal framework to optimize the scale ambiguity problem.Finally,the experimental part is compared with the traditional visual odometry approach and the pure deep learning method to prove the availability of the strategy.Secondly,in specific scenes where point features are sparse,closed-loop detection algorithms based on individual point features perform poorly in detection accuracy.For the sake of the robustness of closed-loop,the characteristics of different information sources are explored,and the closed-loop detection algorithm combining dot-line features is studied.This paper studies the extraction process of image information,introduces the extraction of point and line features,and conducts offline training through public data sets to build a hybrid visual dictionary,which can simultaneously describe environmental contours and structured information.The adaptability of the environment has been significantly improved.Experimental results show that the accuracy of the closed-loop detection algorithm based on the comprehensive characteristics of points and lines has been significantly improved,and it has excellent performance in different scenarios.