| This dissertation designs and realizes a stereo Visual Odometry (VO) system basedon our mobile robot for outdoor environment, and thus lays the foundation for thefurther realization of localization and navigation of our mobile robot in outdoorenvironment.Firstly, the dissertation summarizes the model of VO, and introduces sometechnology involve in VO briefly, such as the local visual feature, motion estimation,RANSAC, etc.Secondly, the dissertation proposed a feature selecting method based on griddingand clustering to improve the accuracy of VO by removing the features located onmoving object regions of the image. The steps of this algorithm are as follows: griddingthe image and extracting the local visual features in every sub-images; selecting thesalient features after clustering all the features with K-means; removing the featureslocated on moving object regions of the image by using RANSAC and epipolarconstraint, and selecting the efficient salient features to estimate the motion. Theexperimental results show that the efficiency and accuracy of VO can be improved byusing the proposed algorithm.Thirdly, the mix accelerating mode with CPU and GPU is proposed after testingthe computation efficiency of feature extraction and matching and motion estimation byusing CPU and GPU. In this mode, GPU is used to accelerate feature extraction andmatching, and CPU is used to perform motion estimation with high efficiency, so thereal-time performance of VO can be improved greatly, which is validated by theexperimental results.Finally, the whole VO system is tested with the track mobile robot developed byourselves, and the results show that the developed VO system can meet the requirementof navigation in outdoor environments for mobile robots. |
PDF Full Text Request