文章目录
基于单应矩阵的视觉里程计算法步骤如下:我们首先来回顾一下,之后本篇博文主要针对ORB特征提取来探索。
The process of homography-based VO can be summed up in the following steps
- Detect keypoints and compute their descriptors from the incoming frame
- Find matches between keypoints in the incoming frame with those in the source frame (e.g. image capture when camera’s pose is C 0 )
- Using RANSAC, compute homography H and also get inliners (matched keypoints that agree with H)
- Decompose H to get (R, t, n) candidates (usually 4).
- Prune candidates that give a negative depth for at least 1 inlier.
- There still may be two candidates, keep the one that has the normal n closer to the current estimation of the plane’s normal vector.
1.ORB 特征提取
The main problem of VO is to estimate the pose of camera from the image.In the classical SLAM model,our keypoints mean landmarks,in visual SLAM,landmarks are the same as image Feature.
角点,边缘,区块A(corner, edge, or block )是图像中特别的点。
角点的提取算法 | |
---|---|
Harris | |
FAST | |
GFTT |
更稳定的局部特征(SIFT,SURF,ORB) ,ORB是性能最好的折中。
repeatability,distinctiveness,efficient,locality
示例原图
我们使用OpenCV来提取下图的ORB关键点和描述子
输出效果
具体的Flag含义可以参考