Rubust Motion Estimation Of Stereo Visual Odometry Based On Attitude Compensation

Robot’s position and attitude are necessary information for navigation. Dead-reckoning is one of the most important means for estimation of position and attitude. But it can not obtain accurate position and attitude information in some special environment. With the rapid development of image processing in recent years, the mobile robot motion estimation using advanced image processing methods has become a hotpot for the robot navigation technology.This paper focuses on the key issues of stereo visual odometry, analyzes the reason of error in detail. LMedS (Least Median of Squares) combining with SFM (Structure From Motion) is applied to get the more precise robot relative position. Attitude compensation is done for the failure of localization caused by pose sudden changes in complex environment, such as uneven road surfaces. The experiments show the correctness and effectiveness of algorithms. The main work is as follows:Firstly, the significance and background of visual odometry are introduced, and then the status research at home and abroad is summarized in detail.Secondly, the theory of stereo vision is given. The common methods of camera stereo calibration and image distortion correction are introduced. The simple Z.Zheng method is realized to calibrate the BB2camera, and the Bought algorithm is realized to image distortion correction.Thirdly, this paper gives the framework of stereo visual odometry. Existing feature extraction and matching algorithms are analyzed, the Harris and SIFT algorithm are compared in experiments. SIFT algorithm is used in the process of visual odometry. The experiments shows that this method has good robustness and real-time.Fourthly, this paper analyzes the existing methods of robust motion estimation and the causes of error. LMedS combines SFM is applied to get the robot relative position. Because the failure of localization caused by pose sudden changes is inevitable in special circumstances, the pose data from AHRS (Attitude Heading Reference System) was added to the visual odomtery algorithm. It ensures the correctness and completeness of the robot trajectory effectively. Experiments showed the method with a high efficiency and a good precision.Finally, conclusion is given with recommendation for future work.