1.概述
表示
注:d11,d12,…可以是法向量,颜色信息
数据来源
激光雷达,RGB-D, CAD models, SFM, 深度图
应用
Robotics, Autonomous driving,Face detection / reconstruction – FaceID Hand pose – Hololens Human pose – Kinect
优点
• 3D information
• Mathematically simple and concise
缺点
• Sparsity
• Irregular – difficulty in neighbor searching
• Lack of texture information
• Un-ordered – difficulty in deep learning
• Rotation equivariance / invariance
2.PCA
Principle Component Analysis
应用
• Dimensionality reduction
• Surface normal estimation
• Canonical orientation
• Keypoint detection
• Feature description
Spectral Theorem
Rayleigh Quotients
利用SVD分析可得到Rayleigh Quotients的结论,以下为证明过程以及SVD的物理意义。U,V代表旋转,特征值代表不同方向的缩放。
3.PCA在点云上的应用
寻找点云主方向![在这里插入图片描述](https://i-blog.csdnimg.cn/blog_migrate/8a4d642325f0c474cb83ce46b396ec2b.png)
Q: What is the most significant principle component?
A: A direction such that the variance of the projected data points on that direction is maximal
Q: How to get the second significant one?
A: Deflation. Remove the most significant component from the data points, i.e., data point minus the projection. Find the most significant component for the deflated data.
Q: How to get the 3 rd one?
A: Repeat the above steps.
证明
算法流程
给一定数量的点云,可以执行如下PCA过程寻找主方向:
寻找点云法向量
Surface normal:The vector perpendicular to the tangent plane of the surface at a point P
应用
• Segmentation / Clustering
• Plane detection
• Point cloud feature for applications like Deep Learning
算法流程![在这里插入图片描述](https://i-blog.csdnimg.cn/blog_migrate/7b4e789231d57fdef08b7cd9d1b180d4.png)
注:点云在最小特征值所对应特征向量的的方向上投影长度最小;曲率暂时不知道怎么应用
考虑噪声
在选择neighbors过程中,如果半径过大,法向量会受到不相关点的影响,如果半径过小,法向量受到噪声影响变大。
解决办法:
• Weighted based on other features
a. Lidar intensity
b. RGB values
• RANSAC
• Deep Learning
以上为学习过程记录,如有错误,请多指教。