|
|
|
Data Processing, Algorithm and Modelling
|
Wavelet-Based filtering the cloud points derived from airborne laser scanner
- Interpolation
The planar interpolation method (Behan, 2000) on the triangulated irregular network (TIN) gives the most accurate
interpolated image. Therefore, this method was utilized in the ALSwave to serve as the first step of processing.
- Wavelet analysis
Wavelet analysis is the core of ALSwave method. A trous algorithm, as mentioned in Section 2, is applied to build
up the multi-resolution framework. The wavelet analysis, therefore, is outlined as follows:
- Input parameter: the number of the resolutions to be analyzed.
- Multi-resolution median filter: the kernel size increases by 2 k + 1, where k=1,2…
- Threshold the median-smoothed images with a threshold of 3 to retain the strong signatures.
- Filter the thresholded images by a trous algorithm (Section 2). Subsequently, differentiate the wavelet-smoothed
images at two consecutive resolutions to generate the detailed images.
Wavelet analysis generates a family of smooth and detailed images, which depict the existence of the objects based
on their size. In wavelet analysis, the only input parameter is the number of resolutions, which is decided by the
operator. It should be known that more resolutions are employed, more information is obtained, and hence longer
time is required to process the algorithm.
- Cluster boundary detection
A simple idea on how to detect the cluster boundaries is illustrated in Figure 2. However, due to the effect of
interpolation and the gap between the laser points, the edges of the objects could not appear vertically. There is a
buffer of edge pixels, named the fuzzy edge pixels. It should be noted that the real location of the edges of the
objects must lie within the buffer zone. The fuzzy edge pixels can be clarified by the additional information from
other data. Unfortunately, this information was not readily available to this study.
Figure 2. Illustration of the detection of the edge pixels
- Selection of the appropriate resolutions
It is an interactive processing where the operator must choose the appropriate resolutions for further processing.
The decision is dictated by the distribution of the objects in the study area and is operating on two limits. While the
lower limit deal with the finer resolution that appears at an acceptable level of noise, the upper limit is related to the
object’s degree of distortion.
- Detection of object points
By a simple spatial relation of fall-into-boundary, it is possible to distinguish between the qualified object points
and the remnants. This processing step is a hybrid method in which the cloud points are grouped and categorized
based on the results obtained from the interpolated images. However, there is an ambiguity along the edges of the
objects. Due to the unpredicted reflectance of laser hit on the objects and the effect of interpolation, several laser
points are classified to the wrong class. Furthermore, the objects such as cars, trees, and poles due to their small size
appeared somehow as noise in the previous processing. Therefore, the laser points belonging to these objects still
exist in the remnants.
- Detection of the fuzzy edge points and clarification of the wrong classified points
The existence of fuzzy edge points has been remarked. These fuzzy edge points must appear within the remaining
set, acquired from the prior processing. It is necessary to filter them out from the bare earth points. It is obvious that
the points belonging to the objects and located at the edges of the objects have a sharp leap in elevation when.compared to the elevation of neighbors. Therefore, elevation threshold is set to classify the fuzzy edge points from
the bare earth points. This processing was performed locally using the Voronoi polygon of the points.
Let OP is the object point set that has been detected and P as the remainders set of points. The fuzzy edge point is a
point that satisfies following condition:
Similarly, the elevation threshold in the Voronoi neighbors could detect the ground laser points that were classified
as the object points.
- Global and local thresholding
The cumulative histogram is checked to find whether any laser points appear with very high elevation on the
ground by the reflection from the complicated objects of the earth surface. Subsequently, a new Voronoi diagram is
generated for the remaining set of points. A slope threshold is carried out iteratively in the local Voronoi neighbor.
As a result, the bare earth points are detected which is ready for the reconstruction of the bare earth surface. Finally,
based on the bare earth point set detected, a triangulated irregular network (TIN) of the bare earth surface is
constructed and the grid-based bare earth surface is generated, subsequently,.
Testing area and the acquired airborne laser data
A typical urban area of 300 m x 300 m in Shinjuku, Tokyo, Japan was selected to test the competence of the
proposed algorithm. There are lots of buildings along with the crowded human activities in this area. The narrow
streets appear in the tiny spaces between the very complex structures of the buildings. In addition, there exist
numerous moving objects on the streets, trees aligned along the streets and buildings. The objects with different
sizes, which are interspersed each other, typified the area. figure 3 illustrates the digital surface model of the test
area.
Figure 3. Digital Surface Model of the tested area
Table 1 shows the parameters of the acquired airborne laser scanner data over the test area, which was provided by
| Operation Altitude |
2700 m |
| Scan Swath Width |
720 m |
| FOV |
16 0 |
| Scan Rate |
19.5 Hz |
| Pulse Rate |
15 KHz |
| Cross Track Spacing |
1.93 m |
| Along Track Spacing |
2.83 m |
| X, Y Positional Accuracy |
0.3 m RMSE absolute |
| Z Positional Accuracy |
0.15 m RMSE absolute |
The approximate laser point density of acquired data is 0.2 point/m 2 . It is quite low for such an application of object
reconstructing in a highly dense urban area. Due to the big gap between the laser points, it was unable to detect the
exact shape of some buildings and hence it was difficult to distinguish trees located close to buildings. The multi-resolution
approach could classify the objects based on their size. Therefore, it is applicable in eliminating the
unwanted trees located nearby the buildings. Both of the filtering, either directly on cloud points or indirectly on
interpolated images, faces unavoidable problem due to this low point density. This is the reason why this study
proposed a hybrid method approach for filtering laser scanner data.
|
|
|
|
|
|
|
