| With the growing popularity of mobile robot, autonomous behavior and decision making is a hot research in the field of mobile robot for the complex outdoor environment. The precondition and primary problem is how to realize the independent positionJnertial navigation combined with GPS is widely adopted outdoor, but GPS is susceptible to multipath error, signal loss and blocking problems, leading to its certain limitation. Location problem is a rising hot spots and research trends in recent years when the signal of GPS is lost or even completely unavailable, and visual odometer technology based on stereo camera has been widely concerned. It is implemented by motion estimation of robot with the captured images by utilizing the coordinate changing of extracted features in image sequence and the cameras model.In this paper, supported by the National Nature Science Foundation of China Project "the localization and environmental modeling of field robots based on set theory for complex terrain environment" (61005092), visual odometry methods based on stereo camera are studied for complex terrain environment outdoor. The specific contents are as follows:The problems of binocular camera and trinocular cameras calibration are studied. Considering the relative position relationship between two cameras, binocular calibration optimization model is established. And the problem of binocular camera calibration is solved. Furthermore a further research of the trinocular camera calibration is done. In view of the redundant information between the three cameras, multi-objective optimization equation with constraints is proposed to improve the calibration precision, which can solve the problem of the trinocular camera calibration.The feature matching and tracking problems in the visual odometry are discussed. Firstly for the images captured by diiferent cameras at the same time, an improved method of feature matching based on SIFT algorithm is presented, and the BBF tree is adopted to speed up the search process in the feature matching. At the same time epipolar constraint and color information etc. are added to effectively eliminate wrong feature matching. The matching accuracy and the performance of real-time are improved in the feature matching. Secondly for the feature tracking in previous and current frame, the correlation of stereo image in time and space is taken into account. The consistency of space position of feature points in two consecutive frames is proposed to filter the tracking points by mistake, which improve matching accuracy and efficiency in the tracking algorithm.Motion estimation algorithm in the visual odometry is researched. Hierarchical motion estimation method that includes the least squares principle combined with RANSAC filtering and two stage bundle adjustment is proposed. Firstly RANSAC filtering methods by means of the iteration is presented to eliminate exterior points in order to obtain more accurate initial estimate. Secondly two stage bundle adjustment that is composed of the iterative optimization for current and previous frame and batch optimization for consecutive frames with the sliding window is presented to optimize the motion estimation results, which can obtain more accurate global positioning results. Lastly the visual information is fused with inertial navigation information by kalman filtering to improve the accuracy, robustness and stability of overall positioning systems.Finally, on the basis of the studies above, MT-FR mobile robot equipped with trinocular camera is used as experimental platform to make the visual odometry positioning experiments in a wide range of outdoor environment. The experiments show the feasibility and effectiveness of the proposed methods in this thesis. |
PDF Full Text Request