Road data in 3-dimensional forms is required for a variety of geospatial applications e.g. road maintenance, transport planning and location-based services. Although airborne lidar can produce dense point clouds from which 3-dimensional road information can be retrieved in detail, lidar data is often incomplete due to the line-of-sight requirement, and therefore a better result of information extraction from lidar data can be achieved if supplement data is used. This paper presents a novel algorithm for road extraction from lidar point clouds and associative vector data. A moving window-based classification technique using various cues determined from the lidar data is applied in a hierarchical way to separate road from other objects. In order to fill the gaps caused by the line-of-sight problem e.g. shadows of trees and buildings, the vector data is used in the refinement process to improve the results by including a buffer, deleting false positives, interpolating and fitting the road-representing points into polylines. To validate the proposed algorithm, four samples of lidar data sets are tested. The test result shows that it is a practical method for road extraction from airborne lidar data for both structured and unstructured lanes.
Feature extraction of urban environments from geospatial data acquired by airborne mapping has been a
significant research topic in photogrammetry and remote sensing for many decades
Urban object extraction is still a sought-after topic, with a focus shifting to the detailed representation of
objects, to using data from new sensors, and to advanced data processing techniques
Most algorithms have exploited height differences, normal vectors, and contextual information to extract
roads from ‘raw’ lidar data. In general, common strategies consist of segmentation and clustering
Because of the shadow in some highly-densed vegetation areas, the results in these areas are not promising.
For the purpose of better accuracy and completeness, some researchers utilized aerial images as well as lidar
data.
Existing geospatial data is a complementary source to serve as a priori geometric knowledge about roads
In this paper, we propose a novel algorithm for road extraction from airborne lidar point clouds and vector data. The proposed method uses a hierarchical moving window that detects road points based on slope, elevation, and intensity values. Because of the highly-densed vegetation areas, the completeness of the results can be challenging to achieve. To refine the initial extraction results, vector data are utilised to improve the accuracy and completeness of the results, including creating a buffer, deleting false positives, filling gaps and fitting into lines. The main difference between existing algorithms and ours is that the vector data is used only in the refinement process to minimise the effect caused by the scale problem and the generalization problem in the vector data. If the vector data is used to partition the neighborhood, there are chances that road points may be left out when a buffer is created to retain lidar points within the buffer since there may be errors in the vector data. Our goal of this work is to develop an algorithm suitable to both structured and unstructured roads.
The proposed method has the following core steps as shown in Fig.1. Firstly, partitioning road points according to x/y coordinates and reordering the points according to y/x coordinates. Secondly, setting the minimum and maximum sizes of moving windows and detecting road points within the given window. Thirdly, enlarging the window size and repeating the detection process until the window size meets the maximum threshold. Lastly, refining the results by vector data, filling the gaps and extracting the centerlines.
Figure 3 An initial process of roads extraction: (a) Part of the profile of lidar points from the transect line in Fig.2, (b) initialise the window size with L0, (c) increase the window size to L1 and repeat the process,(d) increase the window size to L2 and repeat the process
Airborne lidar data contains typically a huge amount of points. Therefore, lidar data possessing is overly
time-consuming and complicated. To reduce the computation time, Vosselman and
After partitioning the lidar data with respect to x and y axes, a moving window is initialized. This process requires the following steps:
Step 1. Set the minimum and maximum window sizes.
Step 2. Set the current window size as the minimum.
Step 3. Detect road points along x and y axes . If the points meet the criteria, increase the classification index of the points by one;
Step 4. Increase the window size by the pre-defined increasing step and repeat Step 3. If the window size is larger than the maximum threshold, do Step 5.
Step 5. Repeat Steps 2-4 until all x and y axes are processed.
Any point whose classification index is above the threshold is marked as the candidate for road points. To distinguish road points from other points, we defined three rules for slope, elevation and height differences, and intensity, as follows. 2.3.1 Rule 1: Slope Since a road is a smooth and continuous surface, the slope of two adjacent road points should be low. Therefore the slope of two adjacent road points should meet the following condition: where is the slope of two adjacent points, and is the maximum slope of the surrounding points.
To some extent, a road is the lowest part of the surrounding points
is the minimum height of the selected road points within the moving window, and
is the self-defined threshold of the height difference.
The distribution of intensity values of ground points should meet the Gaussian Normal Distribution
As seen in Fig. 3(b), if the window size is not properly chosen, some building rooftops can be misclassified as road points. In order to avoid the misclassification, we apply statistical methods to constrain the classification: when the point meets all these criteria with one window size, we add one to its classification index. And the classification index of road points should be larger than the preset threshold; otherwise the point would be marked as an object point.
After the initial roads extraction, some open areas may be misclassified as roads since they present similar characteristics as roads. Further refinement is needed to obtain more accurate results. In this paper, we utilise vector data to refine the initial results. The refinement process includes creating a buffer, deleting false positives, interpolating and fitting to a line, as shown in Fig.4. At first, a buffer is created along each vector segment to constrict the surrounding area and filter out the false positives. Points that fall into the buffer of one road segment are labelled as part of the road segment and will not be taken into account afterwards. If points are outside the buffers of all road segments, they will be labelled as noise and be removed. And then a linear interpolation will be applied along the main direction of each road segment to fill the gaps. To get a smooth surface, fitting is adapted to refine the results. And the centerline is calculated according to the refined results.
(a) Create a buffer
(b) Eliminate false positives (c) Interpolation (d) Fitting the road segments
Four samples of different areas and scenes are chosen for our study as shown in Fig.5 and summarized in Table 1. The first sample is a residential area near the University of New South Wales (UNSW), including 144,255 points. The maximum elevation difference within the dataset is 50.73 m. There are high slopes along the horizontal streets, and in some areas roads are heavily blocked by trees. The second sample is a part of Anzac Parade (an arterial road) with its surrounding residential areas, with 128,847 points. It consists of high-rise buildings, small residential houses, vegetation, structured roads and unstructured roads. In some part of Anzac Parade, it is highly occluded by cars. The third sample is a part of Barker Street (a main road) with its surrounding areas with 124,690 points. Residential houses, trees and roads make up the entire dataset and there is a high slope along the street leading to the fact that in some areas the heights of houses are even lower than that of adjacent roads. The last sample is a part of Botany Street (a main road) with its surrounding areas with 69,064 points. It mainly consists of houses, vegetation and roads. The vector data available is the road network whose resolution is 2 m. Because of the fast development of cities, road geometry can change quickly, which causes a problem to the vector data updating. In this sense, the vector data should comply with the lidar dataset. In this experiment, only the coordinates of starting points and ending points for road segments are available. And the assumption of the vector data is that the road geometry is accurate enough to be used for refinement of the lidar data processing results. 73.60 Figure 5 Four airborne lidar datasets: (a) a residential area, (b) Anzac Parade, (c) Botany Street, (d) Barker
Street.
Based on the proposed method, the minimum window size is set to 15; the maximum window size is set to 35; and the increase step is defined as 5. Then the slope, the height difference, the classification index threshold and the gap between two candidate road segments are also specified as listed in Table 2. The size of window changes gradually from 15 points to 35 points. If a point meets the criterion in all neighborhoods from a window size of 15 to 35, it is regarded as a road point.
Fig.6 shows the road extraction results and respective centerlines of four data samples, which indicates the feasibility of the proposed method. 32.31 61.98
0 58.20 Figure. 6 Results of the extracted road network and respective centerlines: (a) extraction results of Anzac Parade, (b) estimated centerlines of Anzac Parade, (c) extraction results of Botany Street, (d) estimated centerlines of Botany Street, (e) extraction results of Barker Street, (f) estimated centerlines of Barker Street, (g) extraction results of a residential area, (h) estimated centerlines of the residential area.
Completeness = TP/Lr
Correctnes = TP/Le Quality = TP/(Le + FN) (4) (5) (6) where: is the total length of the reference roads; is the total length of the extracted roads; is true positives, namely, the total length of the extracted roads that matches the reference roads; is false negatives, namely, the total length of the undetected roads.
The statistics for each dataset are computed as follows:
The quantitative analysis results are summarized in Table 4, which indicates that Sample 2 shows the worst case. A further inspection indicates that the majority of unmatched roads occurred in two polylines. The main reason is that only starting nodes and ending nodes are available in the vector data and the way points are not present, which obviously contributes to the biases. The majority of undetected roads are the traffic islands and crossings of the two roads. As for traffic islands, although most of the edges are extracted in the initial extraction, they are merged to other road segments when centrelines are extracted. As for the crossings, while they are listed as separated road segments in the reference data, they will be ignored and regarded as one segment in the initial extraction. Roads are important infrastructure and therefore road extraction from lidar data is of great importance in terms of civil engineering and urban planning. In this paper, we presented a novel method for road extraction from airborne lidar data and vector data. The proposed method utilises the local geometry and the radiometric feature of roads to classify road points from the lidar point clouds. To effectively process the data, all points are partitioned with respect to x and y axes and reordered by coordinates. A moving window is initialized to detect the road points. After the initial process, those points whose classification indices are above the threshold are marked as road points..In order to refine the results, vector data is used to eliminate false positives and, in the end, gaps between road segments are filled by a linear interpolation method. The proposed method successfully extracts the roads from airborne lidar data. The quantitative results show that it is a promising method for road extraction, not only suitable for structured roads, but also feasible to unstructured lanes.
Although the presented method is able to extract roads with reasonable completeness, quality and correctness, it is vulnerable to the problem of traffic islands extraction. If not only the starting and ending nodes but also way points are available in the vector data, our a priori geometry knowledge of road segments will help reduce biases in the refinement process. Also, the choice of a window size is critical to the extraction results. If the maximum window size is too small, some building rooftops may fall into the whole window, thus being misclassified as roads. On the other hand, if the minimum window size is too small, it would take too much time for iteration. The aforementioned vulnerability of the proposed algorithm will be the focus of our further research.
Akel, N.A., Kremeike, K., Filin, S., Sester, M., Doytsher, Y., (2005). Dense DTM generalization aided by roads extracted from LIDAR data. In: IAPRS 36 (Part 3/W19), 54–59.
Boyko, A., Funkhouser, T. (2011). Extracting roads from dense point clouds in large scale urban environment.
ISPRS J. Photogrammetry and Remote Sens. 66 (6), S2–S12.
Cheng, L., Zhao, W., Han, P., Shan, J., Liu, Y.X., Li, M.C., (2013). Building region derivation from Lidar data using a reversed iterative mathematic morphological algorithm. Optics Communications (286): 244-250.
Choi, Y.W., Jang, Y.W., Lee, H.J., Cho, G.S., (2007). Heuristic road extraction. In: International Symposium on Information Technology Convergence, IEEE Computer Society.
Clode, S., Kootsookos, P. Rottensteiner, F., (2004(a)). The Automatic Extraction of Roads from Lidar Data.
In: IAPRSIS, Vol. XXXV-B3, pp. 231 – 236.