Abstract: An approach based on least median squares（LMedS） to a stereo-based visual odometry system is presented.Utilizing the scale invariant SITF features in the images as landmarks，the feature matching is implemented with the KD-tree-based nearest search approach to achieve feature matching between stereo images and feature tracking between consecutive frames.Using the 3D reconstruction of the feature，the LMeds method is applied to estimate the information of the robot.The experiments demonstrate the algorithm is robustness in feature matching between different images，the tracking of 3D landmarks and the motion estimation of robot.

Key words: visual odometry, Scale Invariant Feature Transform（SIFT）, Least Median Squares（LMedS）, motion estimation, stereo vision, mobile robot

摘要： 提出了一种基于最小平方中值定理（LMedS）的立体视觉里程计方法。利用图像中尺度不变的SIFT特征点作为路标，基于KD树的最邻近点搜索算法来实现左右图像对特征点的匹配和前后帧间特征点跟踪。通过特征点的三维重建，基于最小平方中值定理估计出机器人的运动距离和方向信息。实验表明该方法在不同图像间匹配、三维路标跟踪和机器人运动估计中具有很强的鲁棒性。

关键词: 视觉里程计, 尺度不变特征变换, 最小平方中值定理, 运动估计, 双目视觉, 移动机器人