Research On Autonomous Navigation Method For Mobile Robots Based On Visual SLAM Optimization Algorithm

With the development of image processing and artificial intelligence in recent years,Simultaneous Localization and Mapping(SLAM)has been widely used in unmanned driving,robot navigation and other fields.The mobile robot explores the surrounding unknown environment through the visual SLAM technology to determine the position and pose and build the environment map,so as to achieve high-precision navigation of the robot.However,there are some problems in visual SLAM,such as high image noise,poor image recognition,and map redundancy.Therefore,this thesis build a mobile robot experimental platform based on visual SLAM theory,a mobile robot autonomous navigation method based on visual SLAM optimization algorithm has been proposed.The visual SLAM system is composed of five modules,and the visual odometry,closed-loop detection and mapping are studied in detail.The main research contents include:(1)This thesis firstly introduces the classical visual SLAM algorithm framework and the corresponding basic theories of visual odometry,closed-loop detection,backend optimization and mapping module.It provides a theoretical basis for the algorithm verification in the following chapters.(2)The mismatching of image features affects the calculation of fundamental matrix,and then lead to poor estimation accuracy of SLAM visual odometry.Aiming at above problems,a visual odometry optimization method based on feature matching is proposed.Firstly,the initial matching set is roughly filtered by minimum distance threshold method,and then the relative transformation relationship between images is calculated by Random Sample Consensus(RANSAC)algorithm.If it conforms to the transformation relationship,it is an interior point.The iteration result with most interior points is the correct matching result.Then homography transformation between images is calculated,the fundamental matrix is calculated by it.The interior points are determined by epipolar geometric constraints and fundamental matrix with most interior points is obtained;Finally,the effects of visual odometry optimization algorithm are verified by TUM data set from two aspects: feature matching and fundamental matrix calculation.The experimental results show that improved feature matching algorithm can effectively remove mismatched feature points while improving the operation efficiency.At the same time,the accuracy of feature point matching is increased by 15.8%;The fundamental matrix estimation algorithm not only improves the calculation accuracy of the fundamental matrix,but also increases the interior point rate by 11.9%.A theoretical basis for improving the accuracy estimation of visual odometry will be provided.(3)Aiming at the problems of low accuracy of visual word bag model and poor image recognition effect in closed-loop detection,a Bo VW algorithm based on improved clustering is proposed.The algorithm firstly calculates the linear decreasing inertia weight and time-varying acceleration coefficient through the image feature fitness value,updates the particle velocity and position.The position of the output optimal particle is the initial cluster center.Then the particles are clustered by the kmeans algorithm until the cluster center does not change,so as to obtain visual words and build visual dictionary tree.The word weight is represented by calculating the Term Frequency-Inverse Document Frequency(TF-IDF),and the similarity score between images is calculated to complete the image recognition.Finally,the optimization model is verified from the two aspects of clustering effect and image classification,the actual closed-loop detection application is carried out based on the optimization model.The results show that the silhouette coefficient value of the clustering algorithm in this model is 0.31 higher than that of K-means,and the visual words are more representative.Compared with traditional Bo VW and FBo W,the image recognition accuracy is increased by 34% and 21%,and image representation ability is stronger;At the same time,the precision and recall rate are improved,which improves the robustness of closed-loop detection.(4)Aiming at the problem of poor mapping accuracy caused by redundant point cloud maps and inability to maintain point cloud features during denoising,an optimization method of dense map based on subsection filtering is proposed.First,the adaptive median filter is used to remove the noise of depth map.Then divide the point cloud based on segmentation idea,set the far distance threshold for direct filtering to filter out outliers.The normal vector of the point cloud at close range is obtained by differential estimation.If the normal direction of the two surfaces is same,the structure remains unchanged,and the redundancy of specific position is removed.Then,the local region fitting function is established by combining the basis function and Gaussian weight function,and the smoothed point cloud data is obtained.It can also further convert dense point cloud map into octree map.Finally,construction effect and accuracy analysis are carried out.In the fr1＿xyz sequence,the experimental results show that the number of point clouds and file size of this algorithm are reduced by 52%compared with RGB-D SLAM algorithm,and the root mean square error values are27% and 12% of RGB-D SLAM and ORB-SLAM2.The algorithm in this thesis purposely reduces redundancy while maintaining the geometric characteristics of the point cloud.The map accuracy is high and the map memory is small.A mobile robot experiment platform relying on RGB-D camera and ROS system is built in the thesis.Based on public data set and the real environment data collected by Kinect v2,the effect of visual SLAM optimization algorithm for mobile robot autonomous navigation is verified.The experimental results show that the proposed algorithm can improve the robustness of SLAM system and facilitate robot positioning and navigation.