融合光流与特征点匹配的单目视觉里程计

针对城市平坦路面准确实时定位的问题,提出将光流跟踪法与特征点匹配进行卡尔曼融合的单目视觉里程计方法.基于平面假设,利用光流跟踪法进行帧间小位移定位,同时利用传统的加速鲁棒特征点（SURF）进行帧间大位移匹配来矫正光流法结果.通过卡尔曼滤波更新机器人的位置和姿态.结果表明,融合算法克服了光流法定位精度差和特征点匹配法处理速度慢的缺点,突出了光流法实时性和特征点匹配定位准确性的优点,该方法能够提供较准确的实时定位输出,并对光照变化和路面纹理较少的情况有一定的鲁棒性.

Monocular vision odometry based on the fusion ofoptical flow and feature points matching

For the problem of real-time precise localization on the urban flat surface, a monocular vision odometry based on the Kalman fusion of optical flow and feature points matching has been proposed. Based on the assumption of flat plane, the method of optical flow tracking was applied for localization between two frames in small movement. Meanwhile, the traditional SURF feature points matching between two frames in long distance was applied for refining the output of the optical flow method. The position and posture of the robot was updated through Kalman filter. The results demonstrate that the fusion algorithm overcomes the shortcomings of poor positioning accuracy of the optical flow and the low processing speed of the feature matching method, highlighting the advantages of real-time performance of optical flow and high accuracy of the feature matching. The fusion algorithm is robust to the circumstances such as illumination change and low road texture, producing a good localization results in real-time.