Research On Location Algorithm Of Robot Based On Monocular Vision

The research on the location algorithm of intelligent robots is an aroused general interest in robotics research and the premise of autonomous navigation of intelligent mobile robots in recent years.In this thesis,a monocular camera is used as a visual sensor to start the research.Firstly,the camera model and Lie group and its corresponding Lie algebra are studied.On this basis,the monocular camera used in the experiment is calibrated and the left perturbation model is expounded,which provide a mathematical model and theoretical basis for subsequent pose analysis and optimization.Secondly,an improved ORB algorithm is designed as the feature extraction algorithm by comparing various feature extraction algorithms and makes the feature points more evenly distributed than the traditional ORB algorithm.And the random sampling consistency algorithm is combined to eliminate the error matching with the feature matching by Brute-Force Matcher.The experimental results show that the algorithm can effectively extract and correctly match the feature points.Thirdly,aiming at the scale uncertainty problem in monocular initialization,a solution is proposed to normalize the initial translation matrix and regard it as a unit.At the same time,the Five-Point method is used to solve the initial essential matrix,which avoids the problem of planar structure degeneracy that would occur in the method of traditional Eight-Point.The simple and effective Perspective-3-Point algorithm is used for the subsequent motion estimation and the Bundle Adjustment is used to optimize the local pose with the left perturbation model based on the Lie algebra to solve the problem.Additionally,the method of keyframe selection is improved to have the robustness and the loop detection method is used for the error accumulation problem when the mobile robot repeatedly reciprocates.And the keyframe is described by using the "words" based on Bag-of-Words which improves the efficiency and accuracy of key frame pairs in the process of loop detection.And the TF-IDF method is used to strictly calculate the similarity of keyframes,which greatly avoids the occurrence of erroneous loop detection.Finally,the overall algorithm is tested with the standard dataset and verified that the maximum error of the motion trajectory with optimization algorithm accounts for about 0.27% of the overall running length and the accuracy is 14.5 times than non-optimized algorithm.And it is also tested in the actual indoor environment and the results proves the practicability and reliability of the algorithm with optimization.