Visual Localization And Environment Modeling For Field Mobile Robots Based On Stereo Camera

With the importance of the application of mobile robots receiving more and more attention,mobile robots application scene is gradually changed from the indoor structure environment to a complex outdoor non-structural environment.At present,GPS combined with inertial navigation is the main method of outdoor localization and navigation.But in some specific environments such as tall skyscrapers,outer space etc.,GPS data is missing or denied.Therefore,how to solve the problem of mobile robots autonomous localization and navigation when GPS data is missing or denied has become a hot topic of research.The visual odometry technology with stereo camera as a sensor and the SLAM method combined with environmental modeling have also received extensive attention.The visual odometry is based on the image information obtained by the mobile robots to calculate the pose information of the robots.Visual SLAM is a method of simultaneous localization and modeling,which solves the problem of localization and modeling at the same time.In this paper,we research the localization and environmental modeling of mobile robots through stereo camera in the outdoor unstructured environment.Specific contents as follows:Aiming at the problem of accurate localization of mobile robots from long distance when GPS data is missing or denied in outdoor environment,a binocular visual odometry system based on ORB feature operator is constructed.Firstly,the ORB feature operator is brought into the binocular visual odometry,and an ORB homogenization feature extraction algorithm based on quadtree partition is proposed due to the non-uniformity of ORB feature extraction.Then,the mismatches of the time-space of the two consecutive frames(4)image is filtered out by the closed loop which is made up by the original feature matching,and the RASAC is used to complete the motion estimation to obtain the exact position transfer matrix.Finally,the EKF fusion inertial navigation data is used to further improve the localization accuracy of the visual odometry,to ensure the real-time and precision of the visual odometry.In order to solve the problem of the rapid reconstruction of outdoor 3D environmental model,a new method of fast environment model reconstruction based on binocular stereo vision is proposed.Firstly,the disparity map is generated by the stereo matching of left and right images,and the 3D point cloud model of single frame is reconstructed by the disparity value and the camera internal and external parameters.Then,the cloud of the overlapping region is optimized according to the relationship between the image pixels and the spatial 3D points in the splicing of the multi-frame point cloud by the position transfer matrix obtained by the visual odometry.Finally,the 3D point cloud is compressed by the octree to obtain the accurate local 3D environmental model.For the SLAM problem of outdoor environment,the position of the robot node is optimized by Graph-base SLAM.First of all,constructing the front-end's graph model of SLAM by the robots pose got by the ORB visual odometry and the key frames' 3D point cloud model.The BoW method is used to generate the word bags model based on the ORB operator to realize the loop closure detection.Then,we apply the back-end's graph optimization to optimize the graph model to obtain the accurate position of the robots,and further improve the localization accuracy of the mobile robots binocular stereo vision.On the basis of the above theoretical research,the MT-FR crawler mobile robots with binocular stereo camera and inertial navigation is used as the experimental platform.A large number of experiments are carried out to verify the feasibility and effectiveness of the proposed method.