Point cloud data denoising

It mainly includes bilateral filtering 、 Curvature flow 、 Density mean drift clustering 、 Noise classification and denoising 、 neural network 、 High density point cloud denoising based on curvature feature hybrid classification Voxel filtering combined with region growth etc.

1、 Bilateral filtering algorithm for point cloud denoising , Bilateral filter is a Gaussian function based on spatial distribution , It can better save the high-frequency information of the target , It smoothes the overall trend of point cloud data , The data points are displaced along the normal direction .
2、 Denoising algorithm based on curvature flow , Every time individual spot Press mirror it Of song rate speed degree Along the the Law towards move dynamic .
Both of the above can make the point cloud model smooth , But at the same time, it will change the coordinates of the point , Lose the texture information of point cloud information . 3、 Density based mean shift clustering Denoising Algorithm , send
Each point gradually approaches the position with the highest local density .
However, cluster noise with high density cannot be filtered
4、 Denoising method of noise classification , Yes Large scale Noise use Radius filtering method and Statistical filtering Method denoising , and Small scale Noise use Bilateral filtering method Denoise .
5、 Adaptive bilateral filtering method for point cloud denoising , The algorithm first establishes ｋ－d Neighborhood , Then take the micro tangent plane at this point as the viewing plane , In this view plane, the spatial variance is realized by using the target scale information σｓ Local adaptive value .（ Medical Science CT The edge of the vertebrae of the image is preserved ）
6、 Using support vector machine and neural network algorithm to denoise High time cost
7、 Using blur Ｃ－ Mean and mean filtering for point cloud denoising Small scale noise and cluster noise cannot be filtered

• 8、 High density point cloud denoising method based on curvature feature hybrid classification .

this paper Focus on the combination of voxel filtering and region growth method
The basic idea ：

1. Gridding point cloud data （ Voxelization ）, The whole can be regarded as a cuboid , Using bounding box compression, it is cut into several equal sized cubes , Make up for deficiencies .

2. Local processing of high-density grid points , Small density grids have both noise and non noise points
   
   

3. Make use of the fact that there is no connection between the noise and the surrounding environment Regional growth method
   Processing steps of regional growth method ：
   （1） Determine a radius a, Find the point with the most points in the grid as the seed point
   （2） The selected seed point is in radius b Combine the surrounding points and use the least square to fit a plane
   （3） radius b If the included angle between the non seed point and the connecting line of the seed point is greater than a threshold , They are classified as noise points, otherwise they are non noise points .
   （4） Take the non noise points in the large grid as new seed points to cycle , Until all points are classified . Use the small grid to grow the global region again .

• It is not explained in the paper （3） How to calculate the medium threshold needs to be studied and the calculation time is long , High complexity , Parameters cannot be taken adaptively