1
|
|
|
""" |
2
|
|
|
Point cloud segmentation using the DBSCAN clustering algorithm. |
3
|
|
|
|
4
|
|
|
DBSCAN - Density-Based Spatial Clustering of Applications with Noise. |
5
|
|
|
Finds core samples of high density and expands clusters from them. |
6
|
|
|
Good for data which contains clusters of similar density. |
7
|
|
|
|
8
|
|
|
See the scikit-learn documentation for reference: |
9
|
|
|
http://scikit-learn.org/stable/modules/generated/sklearn.cluster.DBSCAN.html. |
10
|
|
|
""" |
11
|
1 |
|
import numpy as np |
12
|
1 |
|
from sklearn.cluster import dbscan |
13
|
1 |
|
from patty.utils import extract_mask |
14
|
|
|
|
15
|
|
|
|
16
|
1 |
|
def dbscan_labels(pointcloud, epsilon, minpoints, rgb_weight=0, |
17
|
|
|
algorithm='ball_tree'): |
18
|
|
|
''' |
19
|
|
|
Find an array of point-labels of clusters found by the DBSCAN algorithm. |
20
|
|
|
|
21
|
|
|
Parameters |
22
|
|
|
---------- |
23
|
|
|
pointcloud : pcl.PointCloud |
24
|
|
|
Input pointcloud. |
25
|
|
|
epsilon : float |
26
|
|
|
Neighborhood radius for DBSCAN. |
27
|
|
|
minpoints : integer |
28
|
|
|
Minimum neighborhood density for DBSCAN. |
29
|
|
|
rgb_weight : float, optional |
30
|
|
|
If non-zero, cluster on color information as well as location; |
31
|
|
|
specifies the relative weight of the RGB components to spatial |
32
|
|
|
coordinates in distance computations. |
33
|
|
|
(RGB values have wildly different scales than spatial coordinates.) |
34
|
|
|
|
35
|
|
|
Returns |
36
|
|
|
------- |
37
|
|
|
labels : Sequence |
38
|
|
|
A sequence of labels per point. Label -1 indicates a point does not |
39
|
|
|
belong to any cluster, other labels indicate the cluster number a |
40
|
|
|
point belongs to. |
41
|
|
|
''' |
42
|
|
|
|
43
|
1 |
|
if rgb_weight > 0: |
44
|
1 |
|
X = pointcloud.to_array() |
|
|
|
|
45
|
1 |
|
X[:, 3:] *= rgb_weight |
46
|
|
|
else: |
47
|
1 |
|
X = pointcloud |
|
|
|
|
48
|
|
|
|
49
|
1 |
|
_, labels = dbscan(X, eps=epsilon, min_samples=minpoints, |
50
|
|
|
algorithm=algorithm) |
51
|
1 |
|
return np.asarray(labels) |
52
|
|
|
|
53
|
|
|
|
54
|
1 |
|
def segment_dbscan(pointcloud, epsilon, minpoints, **kwargs): |
55
|
|
|
"""Run the DBSCAN clustering+outlier detection algorithm on pointcloud. |
56
|
|
|
|
57
|
|
|
Parameters |
58
|
|
|
---------- |
59
|
|
|
pointcloud : pcl.PointCloud |
60
|
|
|
Input pointcloud. |
61
|
|
|
epsilon : float |
62
|
|
|
Neighborhood radius for DBSCAN. |
63
|
|
|
minpoints : integer |
64
|
|
|
Minimum neighborhood density for DBSCAN. |
65
|
|
|
**kwargs : keyword arguments, optional |
66
|
|
|
arguments passed to _dbscan_labels |
67
|
|
|
|
68
|
|
|
Returns |
69
|
|
|
------- |
70
|
|
|
clusters : iterable over registered PointCloud |
71
|
|
|
""" |
72
|
1 |
|
labels = dbscan_labels(pointcloud, epsilon, minpoints, **kwargs) |
73
|
|
|
|
74
|
1 |
|
return (extract_mask(pointcloud, labels == label) |
75
|
|
|
for label in np.unique(labels[labels != -1])) |
76
|
|
|
|
77
|
|
|
|
78
|
1 |
|
def get_largest_dbscan_clusters(pointcloud, min_return_fragment=0.7, |
79
|
|
|
epsilon=0.1, minpoints=250, rgb_weight=0): |
80
|
|
|
''' |
81
|
|
|
Finds the largest clusters containing together at least min_return_fragment |
82
|
|
|
of the complete point cloud. In case less points belong to clusters, all |
83
|
|
|
clustered points are returned. |
84
|
|
|
|
85
|
|
|
Parameters |
86
|
|
|
---------- |
87
|
|
|
pointcloud : pcl.PointCloud |
88
|
|
|
Input pointcloud. |
89
|
|
|
min_return_fragment : float |
90
|
|
|
Minimum desired fragment of pointcloud to be returned |
91
|
|
|
epsilon : float |
92
|
|
|
Neighborhood radius for DBSCAN. |
93
|
|
|
minpoints : integer |
94
|
|
|
Minimum neighborhood density for DBSCAN. |
95
|
|
|
rgb_weight : float, optional |
96
|
|
|
If non-zero, cluster on color information as well as location; |
97
|
|
|
specifies the relative weight of the RGB components to spatial |
98
|
|
|
coordinates in distance computations. |
99
|
|
|
(RGB values have wildly different scales than spatial coordinates.) |
100
|
|
|
|
101
|
|
|
Returns |
102
|
|
|
------- |
103
|
|
|
cluster : pcl.PointCloud |
104
|
|
|
Registered pointcloud of the largest cluster found by dbscan. |
105
|
|
|
''' |
106
|
1 |
|
labels = dbscan_labels(pointcloud, epsilon, minpoints, |
107
|
|
|
rgb_weight=rgb_weight).astype(np.int64) |
108
|
1 |
|
selection, selected_count = _get_top_labels(labels, min_return_fragment) |
109
|
|
|
|
110
|
|
|
# No clusters were found |
111
|
1 |
|
if selected_count < min_return_fragment * len(labels): |
112
|
|
|
return extract_mask(pointcloud, np.ones(len(pointcloud), dtype=bool)) |
113
|
|
|
else: |
114
|
1 |
|
mask = [label in selection for label in labels] |
115
|
1 |
|
return extract_mask(pointcloud, mask) |
116
|
|
|
|
117
|
|
|
|
118
|
1 |
|
def _get_top_labels(labels, min_return_fragment): |
119
|
|
|
"""Return labels of the smallest set of clusters that contain at least |
120
|
|
|
min_return_fragment of the points (or everything).""" |
121
|
|
|
|
122
|
|
|
# +1 to make bincount happy, [1:] to get rid of outliers. |
123
|
1 |
|
bins = np.bincount(labels + 1)[1:] |
124
|
1 |
|
labelbinpairs = sorted(enumerate(bins), key=lambda x: x[1]) |
125
|
|
|
|
126
|
1 |
|
total = len(labels) |
127
|
1 |
|
minimum = min_return_fragment * total |
128
|
1 |
|
selected = [] |
129
|
1 |
|
selected_count = 0 |
130
|
1 |
|
while selected_count < minimum and len(labelbinpairs) > 0: |
131
|
1 |
|
label, count = labelbinpairs.pop() |
132
|
1 |
|
selected.append(label) |
133
|
1 |
|
selected_count += count |
134
|
|
|
return selected, selected_count |
135
|
|
|
|
This check looks for invalid names for a range of different identifiers.
You can set regular expressions to which the identifiers must conform if the defaults do not match your requirements.
If your project includes a Pylint configuration file, the settings contained in that file take precedence.
To find out more about Pylint, please refer to their site.