Research On Key Issues Of 3D V-SLAM And Motion Control For A Mobile Robot

Autonomous navigation is the process that a mobile robot can independently explore the surroundings and estimate poses of itself by sensors,meanwhile the mobile robot can autonomously perform collision-free motion from the starting position to the target position without human intervention.The autonomous navigation technology of a mobile robot involves Simultaneous Localization and Mapping(SLAM),path planning,and motion control.SLAM is the premise and key of a mobile robot's autonomous navigation.Motion control is the foundation and guarantee for a mobile robot to accomplish various tasks.This dissertation focuses on the issues of three-dimensional visual SLAM(3D V-SLAM)and motion control for the autonomous navigation of a mobile robot.Recently,Kinect-based 3D V-SLAM method has turned out to be the mainstream technique of SLAM for a mobile robot,which is composed of front-end thread and back-end thread.In the front-end thread,the local environment model is updated and the pose of the mobile robot is estimated by feature extraction and matching and point-cloud registration.Then,in the back-end thread,the result obtained by the front-end thread is further optimized globally.Therefore,the accuracy of feature extraction and matching and point-cloud registration has significant influence on the performance of 3D V-SLAM for a mobile robot.From the perspective of control objectives,motion control can be classified into point stabilization control,trajectory tracking,and path following.Usually,in the process of the autonomous navigation,the change of a mobile robot's pose is large in the presence of external environmental disturbance.In this case,large viewpoint changes might occur in the images and point-clouds obtained by Kinect,as well as big tracking errors in the motion control.Hence,how to guarantee the effectiveness of the feature extraction and matching and point-cloud registration,the accuracy and rapidity of trajectory tracking and path following are crucial for a mobile robot to realize the high-performance autonomous navigation.In this dissertation,the above key issues are studied and the main contents are as follows:1.Considering the problem of feature extract and matching under a large viewpoint change,a fast and fully affine-invariant algorithm A-ORB is proposed.The algorithm is inspired by the idea of simulating affine transformation used in Affine-SIFT(ASIFT)algorithm,which guarantees the ability of resisting viewpoint changes.Oriented FAST and Rotation BRIEF(ORB)algorithm is used to extract features from the stimulated images,which improves the rapidity of the feature extraction process.Hamming distance is used to replace Euclidean distance to shorten the matching time.Experimental results demonstrate that A-ORB algorithm not only inherits the capability of fully affine-invariance of ASIFT algorithm,but also maintain the rapidity of ORB algorithm.2.Considering the time consuming problem in the process of feature description and the redundancy steps in the process of feature matching in A-ORB algorithm,a faster and fully affine-invariant algorithm AFREAK is put forward.In this algorithm,Fast Retina Key-point(FREAK)descriptor is used to solve the time consuming problem.A new matching strategy is designed to simplify the matching steps.Experimental results verify that AFREAK algorithm can improve the rapidity obviously and preserve the fully affine-invariance.In addition,an image stitching method based on AFREAK algorithm is developed to solve the time consuming problem when the image views are too different,and the rapidity of AFREAK algorithm is further demonstrated.3.Considering the point-cloud registration problem when point-clouds have a large viewpoint change,a point-cloud registration method based on ASIFT is proposed.In this method,affine-invariant features and matches are obtained by using ASIFT algorithm,and Optimal Random Sample Consensus(ORSA)algorithm is used to eliminate outlier matches.3D feature point-clouds are generated by inlier matches and corresponding depth information,unit quaternion method is adopted for computing initial transformation.Iterative Closest Points(ICP)algorithm based on features is used to achieve precise point-cloud registration.Experimental results indicate that the proposed method can achieve point-cloud registration with high precision in the event of large viewpoint changes in point-clouds.4.To realize a high-performance 3D V-SLAM for a mobile robot in the presence of large viewpoint change,a 3D V-SLAM system based on the proposed AFREAK algorithm,point-cloud registration method,and optimal algorithms is developed.The offline(based on standard benchmark data sets)and the online(in real scenes)experimental results show that the developed 3D V-SLAM system can accurately reconstruct the environment,detect the loop closures and estimate poses of the mobile robot.5.To deal with both velocity-jump and velocity-tracking problems in back-stepping trajectory tracking control for a mobile robot in the presence of large tracking errors,a trajectory tracking control system based on biological membrane voltage model and back-stepping sliding mode control technique is developed.Simulation results verify that the developed control system can effectively solve the problem of velocity-jump,realize the accurate velocity-tracking,and trajectory tracking.6.Considering the problem of long convergence time occurred in the path following control of a mobile robot adopting first-order dynamic sliding mode control approach and exponential reaching law,a fast path following control system by employing the backstepping sliding model control technique and double power reaching law is proposed.Simulations and practical experiments are carried out and validate the rapidity and accuracy of the proposed control system.