segmentation-based classification for 3D points clouds in the road environments
section 1: introduction
Listing MLS(mobile lidar scaning) and the main contributions of our work
two questions
- noise, occlusion caused by obstructions
- density cariation caused by different distance
our approach
step1: original segmentation (pairwise linkage and two-step post-processing method)
step2: extract a ste of features of each segment for training and testing by using a classifier.
step3: refine the classifier by using the contextual constraints via graph cuts
section 2: related works
two main categories of 3D points classification
- the point-based classification
- the segmentation-based classification
point-based methods
- Aparajithan: label a single point as either a ground one or a non-ground one.
- GUO :multiole returns and echoes of laseer pulse can be discriminant features in the neighbouhood point cloud calssification.
- the methods od neighborhood selection: spherical,cylindrical,voxel,k-nearest neighbourhood and conbination of multi-scale and various neighborhood.
- questions: computational complexity, fail in some complex point cloud classification conditions due to the limitation of features extracted at the point scale.
segmentation-based classification
- surface discontinuities(normals, curvatures,height differences are wildly used in the smoothly connected areas)
- graph cuts and markov filed are used to generate the smoother segmentant than the traditional region-growing methods, this method is frist to construct a weighted graph where each edge weight cost represents the similarity of the corresponding segments,and then find the minimum solution in the graph.
- questions:under-segmentation(all the segments are assigned to be the same lable)-> hierarchical segmentatonmethod, this method produced over-segmented results and then use the post-processing strages used to merge the over-segmented parts.
classification algorithm
- hierarchical tress
- machine learning methods (SVM to distinguish basic categories,Random Forests ,AdaBoost algorithm)
- conditional random fields(CRF):the description of CRF :
section3: our approach
3D point cloud segmentation
P-linkage Based Segmentation
- Normal estimation:the KNN based methods is employed to find the neighborhood of each data point and estimate the normal by PCA. By using PCA,we finally obtain the normal n(pi), flatness lamda(pi),and the consistent Set CS(pi).
- linkage building: if pi has the CNP(pi),then a pairwise linkage between CNP(pi) is created and recorded into a lookup table, otherwise, pi is considered as a cluster center and inteter into the list of cluster centers .
- Slice Creating: surface slice Sp.
- Slice merging.
two post-processing steps
- step one :group the broken cars,trees and curbs into the whole ones by using the connected components analysis.
- (Scandidate—>get the geometric features(scatter,liner,planar) of the point p by using PCA—>reigon growing,)
- step two: try to merge the co-liner segments.
segmentation-Based classification
feature extraction
predefined features
-
orientation:normals
-
heights:find the grounds—>calculate the average height of the segments—>slove the clost objects by using MSE
-
Geometrical shapes: projections
-
point types:scatter, horizontal, slope, vertical
-
densities:
calssifiers
- SVM(support vector machine)
- RF(random forest)
- ELM(extreme learning machine)
optimization via graph cuts
results
datasets
seven sets.
segmentation
s1,s2,s3 contain the noise —>test the robustness---->merge the near segment