|
| 基于点线特征融合的立体视觉里程计 |
| Stereo visual odometer using point and line features |
| 投稿时间：2020-10-21 |
| DOI：10.3969/j.issn.1005-5630.2021.04.003 |
| 中文关键词: 视觉里程计 位姿估计 点线特征 重投影误差 |
| 英文关键词:visual odometer pose estimation point-line features re-projection error |
| 基金项目:国家自然科学基金（61374197）、上海市自然科学基金（20ZR1437900） |
|
| 摘要点击次数: 3693 |
| 全文下载次数: 2077 |
| 中文摘要: |
| 大多数视觉里程计通过跟踪图像序列中点特征几何位置的变化实现对相机位姿的估计。线是点的集合，相对于离散的点，帧间线特征的位置变化更具有显著性，因而有利于提高特征检测跟踪的鲁棒性。另外，在一些弱纹理场景中，点特征不够丰富，作为对点特征的补充，提出了一种融合点线特征的立体视觉里程计算法。构建新颖的点线重投影误差模型作为目标函数求解旋转矩阵和平移向量。模型中，使用Huber核函数减小特征误匹配对优化过程的影响。选取ORB算子检测点特征，LSD算子检测线特征，匹配时施加恒速约束、环形匹配、Bucketing约束和外观几何约束，提高特征匹配的速度及精度。采用公共数据集KITTI和EuRoC对算法进行评测，实验结果表明，该算法在多种场景中的鲁棒性能，相较于其他具有代表性的视觉里程计算法，在精度方面有提升。 |
| 英文摘要: |
| Majority of existing visual odometers estimate the pose of camera through tracking point features and examining changes of their geometric position between consecutive frames. Comparing to discrete points, position of a line changes more significantly over frames. Therefore, detection and tracking of lines is more robust than that of points. In addition, line features are more abundant in low-textured scenes. As a complementation of point features, this paper proposes a novel stereo visual odometry that combines point and line features. A point-line re-projection error model is constructed as the objective function to estimate the rotation matrix and the translation vector. Huber kernel function is used in the model to reduce the impact of mismatching features on the optimization computation. ORB is used here to detect points and LSD is used to detect line segments. Speed smoothness constraint, loop matching, bucketing constraint and geometrical constraint are imposed on feature matching process to improve the matching speed and accuracy. Experiments have been conducted on public KITTI and EuRoC datasets, and the results demonstrate that our algorithm can work robustly in a variety of scenarios and has better performances in comparison with other representative visual odometry. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
