Patty Analytics¶
Patty Analytics aims to register pointclouds that were generated from photos or video to an absolute position, scale and orientation.
Pointclouds generated from photos are generally messy; they have holes and floating unidentified objects. In our scripts we assume to have the following information: a map (drivemap) which has an extremely low resolution but has good absolute coordinates; a footprint polygon denoting more or less the latitude and longitude and area of the object (x and y coordinates). Finally, we have the high-resolution pointcloud of the object. By the nature of creating this pointcloud, it is densest at the object, since the photos usually center on this object. In some cases, there are also camera positions available, relative to the object.
Registration¶
The registration.py script is the main registration algorithm. This script can be run as follows:
``` Usage:
registration.py [-h] [-d <sample>] [-u <upfile>] <source> <drivemap> <footprint> <output>
- Positional arguments:
- source Source LAS file drivemap Target LAS file to map source to footprint Footprint for the source LAS file output file to write output LAS to
Example run:
`
python scripts/registration.py data/SOURCE/SITE_X.las data/DRIVEMAP/X.las data/FOOTPRINT/X.footprint.csv data/OUTPUT/SITE_X.las -d 100000
`
Initialization¶
The registration script applies a dbscan clustering algorithm before anything else to filter out noise and end up with the densest parts of the pointcloud. This is the slowest operation. The algorithm includes at least 70% of the pointcloud. To speed this operation up and make its runtime predictable, the pointcloud can be subsampled to a given number of points before doing the clustering algorithm. A uniform probability distribution is used for this.
Orientation¶
The registration script does not yet give a definite orientation. It assumes that objects stand on the ground, and that the footprint represents some part of the ground. In turn we assume that both the pointcloud and drivemap are more or less flat at this location. The first step is then to detect the boundary of the object, resulting in a line-like subset of the pointcloud, usually representing its demarcation at the ground and holes that have formed during pointcloud generation. Then, a small margin around the footprint is cut out of the drivemap. The orientation is then determined by aligning the two principal axes of the pointclouds boundary to the drivemap cutout. The principal axes are found by taking the first two components of a PCA of the drivemap and the cutout. With these operations, it is still not possible to distinguish up from down and front from back.
The next step is to, when available, use the camara positions to determine an approximate UP orientation. This orientation is used to determine up from down. Unfortunately this still does not determine front from back.
Scale¶
In our project, red-white meter sticks of 80 cm were placed next to the objects, so this is a natural way to determine the object’s scale. These sticks have two red and two white parts, each 20 cm long. The red parts are rare in many natural environments, so we first take all red points (in HSV color space: H > 0.9 and S > 0.5). Then, the remaining points are clustered with the dbscan algorithm. The length of each cluster (each red stick part) is determined by applying principal component analysis to find the longest axis and measuring this axis. If three or more clusters have about the same length and high number of points, we assign a high confidence level to the stick scale. The fewer points we have and the fewer conforming clusters, the lower our confidence that this scale is the correct one.
If these meter sticks are not available or not well represented, we determine the scale by comparing the size of the oriented boundary of the object to the size of the footprint, and take that as the scale.
Position¶
Once scaled and oriented, the position of the pointcloud is determined by shifting the pointcloud boundary to the location of the footprint. Its height is determined by the height of the cutout of the drivemap around the footprint.
Future work¶
Once the pointcloud is placed at its approximate position, a few measures could be taken to improve its position. One way is to use ICP or GICP from the PCL library. Since the density of the drivemap and the pointcloud differ greatly, the poincloud probably has to be subsampled first. A complicating factor is that the drivemap may have a grid-like topology because of its low density whereas the pointcloud does not, which is likely to throw an ICP algorithm off.
Additionally, a measure of fit could be used (nearest point cumulative distance comes to mind) to try a few common operations. The following operations come to mind: rotating the front to the back; scaling between 50% and 150%; and slightly moving the object to each side.
The Python-PCL library currently does not support point normals. If it did, it would give additional tools to determine rotation.
Pointcloud utilities¶
In addition to registration, this package has a few utilities for dealing with pointclouds. In transform.py, pointclouds can be rotated, scaled and translated. With convert.py, a pointcloud can be converted from one format (of PCD, PLY or LAS) to another. The Spatial Reference System (SRS) of a LAS file can be set with las_set_srs.py. Point normals are not preserved in these operations. In the tests/helpers.py file, synthetic pointclouds can be generated for means of testing.
Contents: