Research on Urban Road Network Extraction Based on Web Map API Hierarchical Rasterization and Improved Thinning Algorithm

Abstract

Most existing research on the vector road network is based on GPS trajectory travel information extraction, and urban GPS trajectory data are large and difficult to obtain. Based on this, this study proposes a road network extraction method based on network map API and designs a vector road network based on an improved image-processing algorithm using trajectory data. Firstly, a large number of trajectory data are processed by hierarchical rasterization. The trajectory points of the regional OD matrix are obtained by using the map API interface to generate the trajectory. Then, the image expansion processing is performed on the road network raster image to complete the information loss problem. The improved Zhang–Suen refinement algorithm is used to refine the idea to obtain the road center line, and the vector road network in the study area is obtained. Finally, taking the Harbin City of Heilongjiang Province as an example, compared with the road network of the network map, it has been demonstrated that using this technology may improve the traveler experience and the sustainability of urban traffic flow while reducing the number of manual procedures required, performing online incremental rapid change detection, and updating the present road network at a cheaper cost.

1. Introduction

With the rapid development of urban road construction, people’s demand for travel is increasing, but road network information is updated slowly. The vector road network has high application value in land space planning, intelligent transportation, urban construction, and disaster relief [1,2]. The traditional road network acquisition methods are manual surveying and mapping, geographic information collection, and satellite image processing. The update technology cycle is long, and the cost is high [3,4,5]. With the rapid development of information technology, the new road network generation algorithm has become a research topic for many scholars. The most common method is to generate a road network based on vehicle GPS trajectory data or walking GPS trajectory data [6,7,8]. Road network extraction based on GPS trajectory has the advantages of high precision, low acquisition cost, comprehensive data source, and strong real-time performance for urban roads when compared to the conventional methods of surveying, mapping, and image interpretation to obtain digital map road network data, but the road network data are often small-scale. However, the magnitude of GPS trajectory in an urban area is significant, and generating a road network using GPS trajectory is challenging. The Open Street Map (OSM) project is a free, open-source, editable map service for the online public.

Nevertheless, the OSM vector map obtained from the Internet has serious road missing problems in some underdeveloped areas [9,10,11,12,13,14]. In this study, considering the integrity of the urban road network, a method of trajectory generation and road network extraction based on map API is proposed, which can obtain the geometric information of the regional road network and make it possible to construct or update the road map in real time. In [15], for instance, relevant road information (such as geometric features, speed, latitude and longitude, direction angle, etc.) is extracted and updated from the trajectory data in the vicinity of the road segment; in [16], the K-means algorithm is used to cluster the trajectory points, and the spline curve is introduced to fit the road turning area; and in [17], the straight line and arc in the mathematical model are used to describe the GPS trajectory. According to the positioning point and the driving direction, a large number of GPS trajectory points are clustered and connected to form a trajectory chain, and the road center line is fitted by the nonlinear least squares method. In [18], the trajectory points are classified, the spatial and semantic relations are judged, and the trajectory points are merged into the candidate road network; in [19], a general map generation method is proposed based on trajectory merging, and the geometric features of the road are inferred by artificial intelligence algorithm. The clustering algorithm is used to incrementally generate the road network. The authors of [20] extracted the central axis of the road based on the constrained Delaunay triangulation to establish the street road network; Ref. [8] introduces the constrained Delaunay triangulation and its dual Voronoi diagram geometric model to extract the road boundary from the vehicle trajectory line. The authors of [21] rasterized the vehicle GPS trajectory data, merged the discrete data points based on mathematical morphology, and finally used the SPTA algorithm to refine the road network; Ref. [22] created a buffer for the vehicle trajectory data and then used the cross-number refinement algorithm to refine the taxi trajectory data to build the road network. In general, the grid refinement method converts a large number of vector trajectory data points into image operations, which improves computational efficiency.

2. Materials and Methods

Firstly, the research area is rasterized, and the path trajectory points are obtained through the map navigation service to generate the trajectory. The GPS trajectory data used in this study are from Didi Chuxing Technology Co., Ltd., Beijing, China, The Didi Chuxing GAIA Initiative. These GPS data are collected in the vehicle GPS equipment of the Didi Chuxing platform vehicle. As shown in

Table 1. Partial original GPS trajectory data.

Vehicle_id	Order_id	Utc	Longi	Lati
7n6dhjnad_Et0hsuh k9ee7fjf2DE-ovu	gjhdafj6h@tz2rAy hnga7lkgh2ytaruA	20220512T152036+08	126.648304	45.753281
9mde7Jj9aJEv6gsrh fk5e7d982EwanBni	dlh48kcie6Dticiyyg 17i7cgb69vubcpwt	20220513T081325+08	126.637527	45.756314
icg876n8h7vq8hxq gbj776hcb9Dscmyo	8hk99lh739qc]lkpz 9n9e7jk6cOAqleDq	20220513T105703+08	126.590429	45.730337
9mde7f)9a_JEv6gsrh fk5e7d982EyvanBm	dlh48kcie6Diiciyyg 17i7cgb69vubqywt	20220515T142516+08	126.527859	45.814551
6nehde97f5Br. oD fbhcffef5btE.jAp	efa5c6bf93wB8gv wihl3hehab-rr. hBv	20220516T174506+08	126.678457	45.794172
ailbh7ly93DF0iux aohe4gkl42zxllxv	gbdf38nid3DiiclDtb 9ih7ari6h.AE9duz	20220517T201538+08	126.692157	45.772399

, the data fields include the driver number (vehicle_id), order number (order_id), world time (utc), longitude (longi), and latitude (lati). Secondly, the accepted course is rasterized, and then the vector road network is constructed based on rasterization. The technical route of trajectory generation and road network extraction using the map API are shown in Figure 1.

2.1. Acquisition and Hierarchical Rasterization of Trajectory Data

This study proposes a rasterization method of a hierarchical index. In the face of massive trajectory data, to improve the efficiency of rasterization, the selected areas where the trajectory points fall into are graded, and the direction, speed, and time indexes are established to preserve and utilize the information in the grid, as shown in Figure 2.

According to the size of the selected area, the selected area is divided into three levels, corresponding to Div, Grid, and Cell in Figure 2. Div is the partitioning placement method used in the stacked stylesheet. This research refers to the study area; Grid describes an image that has been discretized on a two-dimensional plane and refers to brightness. The grid picture can be viewed as a matrix. Any element in the matrix can represent a point in the image, and the matching value can be used to represent the point’s grayscale. Pixels are the units that make up the digital matrix. The purpose of this classification is to categorize the research area’s level of gray; a Cell is the smallest record point unit, for the extraction of the road centerline, the size of the Cell is the same as that of the vehicle, about 5 m × 5 m, and then the direction index, time index, and speed index are established for Cell, as shown in Figure 2. Among them, the direction index divides the direction into eight different ranges, in which 0~7 represents different directions, the time index is in hours, 0~23 represents different periods, and the speed index divides the speed into (0~9) 10 different ranges according to the speed of the trajectory point.

The trajectory data is processed in days. For each trajectory point in the selected area Div, the Grid to which it belongs is first recorded. Then the Cell is found to which the corresponding Grid belongs and the trajectory points in the corresponding Cell are calculated. The trajectory points’ direction, time, and speed information are also counted within the respective ranges of the Cell direction, time, and speed indexes. This is used to record the number of trajectory points in the unit and count the number of trajectory points in different directions, time periods, and speed ranges in the unit.

The multi-day trajectory data can complement some of the missing road information, so the number of different trajectory points in the same Cell for multiple days and the statistics in different index ranges are added. Using the trajectory point speed information in the Cell speed index and the trajectory point azimuth information in the azimuth index, the trajectory points with high- and low-speed ranges or unreasonable directions in the Cell statistics are removed. If there are trajectory points in Cell, the pixel value is 0 if there are none, and 255 otherwise. This allows us to acquire the grid road network image.

Then, the research region is rasterized by fusing changes in the vehicle speed, urban latitude and longitude, and GPS data-collection period. To create grid center point C and represent the research region, a grid cell of 50 m × 50 m was built. Symbols can be simplified to:

$$\mathrm{A}=\{C_{o1},C_{o2},C_{o3,}\dots ,C_{on}\}$$

(1)

where A represents the study area, C_o represents the center point of the grid cell, and n represents the total number of grid points in the study area. To obtain a more accurate road network, the resolution of the grid unit can be set to a smaller value, but it is not the smaller the better, as a 1 m × 1 m grid unit cannot distinguish whether the collected valid data or positioning drift error are caused by abnormal data points. The purpose of modeling is to obtain the study area’s starting- and endpoint matrix (OD matrix). Taking point C_o as the starting point and its two adjacent units in the southeast and northwest as the endpoint, the OD matrix of point C_o is constructed, which is expressed as: $\left[\begin{array}{c}{C}_o\\ C_{d1}\\ C_{d2}\\ C_{d3}\\ C_{d4}\end{array}\right]$. Then, the OD matrix of the entire study area can be described as: $\left[\begin{array}{cccc}{C}_{o1}& C_{o2}& C_{o3}& C_{on}\\ C_{d1}& C_{d1}& C_{d1}& C_{d1}\\ C_{d2}& C_{d2}& C_{d2}& C_{d2}\\ C_{d3}& C_{d3}& C_{d3}& C_{d3}\\ C_{d4}& C_{d4}& C_{d4}& C_{d4}\end{array}\right]$. where C_o is the center point of the grid unit, which is called the starting point, and n is the total number of grid points in the study area. C_d1, C_d2, C_d3, and C_d4 represent the grid center points of two adjacent units in the southeast–northwest direction corresponding to the starting point, which is called the endpoint here (see Figure 3).

In this study, the working image’s center point is determined using an estimating technique based on density weight. To achieve this, the grid density is utilized as the weight, and the weighted average algorithm is used to obtain the latitude and longitude of each point’s equivalent center. Symbols can be simplified to:

$$lon={\displaystyle \sum _{k=1}^n\frac{p_k}{p_{sum}}}\times lo{n}_k$$

(2)

$$lat={\displaystyle \sum _{k=1}^n\frac{p_k}{p_{sum}}}\times la{t}_k$$

(3)

$$p_{sum}={\displaystyle \sum _{i=1}^n{p}_i}$$

(4)

Here, n denotes the total number of grids in the research region, p_sum the overall density of the n grids, p_k the density of the k th grid, and (lon_k, lat_k) the latitude and longitude of the k th grid’s center of mass. The working area’s center may be found at the computed (lon, lat) coordinates.

The map navigation service obtains the path points based on the OD matrix. If the navigation trajectory between any two points is traversed, an n × n Cartesian product (n refers to the number of grid points in the study area) will be generated, which increases the difficulty of the operation, and a large number of repeated paths will occur. If the navigation path of each point and its four neighborhoods is traversed, the complexity of the operation is reduced from n × n to n × 4. Even while it makes calculations easier, some secondary roads or branches have fewer cars going through, which means the findings will result in a discontinuous trajectory, making it challenging to maintain the acquisition of the road network. According to the efficiency of the comprehensive operation and the rationality of the trajectory, the grid center point of the two adjacent units in the southeast and northwest corresponding to the starting point is selected as the endpoint. The trajectory point set of the entire region is obtained by traversing the entire region’s target point and its corresponding endpoint. The path trajectory points include the longitude, latitude, and road type (

Table 2. Navigation map parameter diagram.

Start_Lng	Start_Lat	End_Lng	End_Lat	Road_Types
126.641046	45.723925	126.637141	45.725064	main road
126.637141	45.725064	126.645981	45.737825	main road
126.645981	45.737825	126.630618	45.746900	secondary road
126.630618	45.746900	126.643020	45.757201	main road
126.643020	45.757201	126.634909	45.761033	main road

). Symbols can be simplified to:

$$p=(x,y,i)$$

(5)

where p represents the trajectory point, x the longitude, y the latitude, and i the navigation command for the trajectory point.

Finally, the trajectory point generates a trajectory. The map navigation service returns a set of navigation trajectory points, which needs to be converted into a trajectory. According to the set of trajectory points belonging to the same command, a trajectory l is formed in chronological order, and all trajectory points are processed in the same way to obtain the network trajectory T of the whole study area.

The trajectories obtained by the map service will overlap substantially in different parts of the same road. This study draws on the grid clustering method to perform trajectory aggregation processing [23]. In practical applications, due to a large amount of GPS data processing, taking into account the real-time performance of the algorithm, this study does not use the original GPS trajectory data but is based on the grid unit equivalent centroid coordinates, using the least squares method to fit the road curve. That is, for a pile of discrete data points, find a curve y to minimize the sum of deviations from all discrete points. The polynomial curve fitting formula is:

$$y=a_0+a_{1}x+\dots +a_k{x}^k$$

(6)

where a_k is the coefficient of the kth factor, k is the polynomial degree, the fitting result is a straight line when k = 1, and the fitting result is a curve when k ≥ 2. The binarization method processes the trajectory data into a black-and-white binary image.

Hierarchical rasterization has two advantages: (1) The entire region is divided into different levels, and only the trajectory points fall into the previous level of Cell to perform the calculation; the entire selected area is not used to perform the calculation, so the speed compared to the traditional rasterization is improved. (2) Information such as the speed and direction of the track point is reserved in the grid point, and it can be used to judge whether the track point is reasonable or not, and the noise point can be removed to improve the accuracy of the grid road network image. However, there are two main points of hierarchical rasterization: (1) According to the selected area, the larger the area, the greater the amount of grading, and the minor area level must be Cell; the Grid between Cell and Div can be divided into multi-level according to the area. (2) The Cell size needs to be controlled; too much detailed information is easy to lose, and too small a neighborhood lane will increase the difficulty of centerline extraction.

2.2. Application of Fast Expansion and Improved Thinning Algorithm in Road Network Extraction

In the grid road network image obtained, there are cases wherein missing data lead to road disconnection and adjacent lanes. The separate road sections need to be connected, and the road centerline extraction needs to be merged with adjacent lanes. The expansion algorithm can expand the road section to the surrounding area, connect the road section, and merge the lanes. A simple and fast expansion algorithm is to scan the pixels in the image in turn, find the pixels with a gray value of 255, analyze the eight pixels around the pixel, and set the gray pixel value to 0 only if there are pixels with a gray value of 0 in the eight pixels.

After the expansion algorithm is processed, the road network image shows the characteristics of the road network. However, the road section is a multi-pixel width, and the extraction of the road section is more complex. The thinning algorithm can maintain the topological structure of the black part of the image. After processing, it shows a single pixel width. The skeleton line in the image is the center line of the road. This study uses an improved Zhang–Suen thinning algorithm. Binarizing the expanded image, for the point with road information (pixel value is 0), its pixel value is set to 1, called the target point, and for the non-road information point (pixel value is 255), its pixel value is set to 0, called the background point; a simple binary image can thus be obtained.

The process of improving the Zhang–Suen thinning algorithm is as follows:

(1)

Find eight neighborhoods centered on the target point, and the center point is P₁. The eight points in the neighborhood are recorded as P₂, P₃,..., P₉, and P₂ on P₁. Mark pixels that meet the following conditions:

①

2 ≤ N (P₁) ≤ 6

②

S (P₁) = 1 or B (P₁) ∈{ 65, 5, 20, 80, 13, 22, 52, 133, 141, 54}

③

P₂ × P₄ × P₈ = 0 or S (P₂) ! = 1

④

P₂ × P₄ × P₆ = 0 or S (P₄) ! = 1

(2)

Remove all points that meet the mark, and assign the gray value of the target point to 0 and the gray value of the background point to 255.

Where N (P₁) is the number of non-zero neighbors in the 8-neighborhood of P, and S (P₁) is the number of changes from 0 to 1, starting from any point in the 8-neighborhood around P, in turn, along the counterclockwise direction. B (P₁) is the binary code S of the P₁ neighborhood point. The calculation formula is shown in the formula below, and the value of P_i is its pixel value.

$$\mathrm{S}=\underset{n=2}{\overset{9}{\Sigma }}(P_n\times 2^{n-2})$$

(7)

The improved Zhang–Suen thinning algorithm uses a 5 × 5 neighborhood and changes the original Zhang–Suen algorithm thinning conditions based on increasing the neighborhood range. The effects of three thinning algorithms for an intersection are shown in Figure 4. The light color part is the effect of the intersection after expansion, and the dark part is the effect after thinning. Figure 4a shows the Zhang–Suen thinning algorithm, which is not a single pixel width after thinning, and the effect at the intersection of the road section is poor. Figure 4b is the improved thinning algorithm in [8], only changing Figure 4a to a single pixel width; Figure 4c is an improved Zhang–Suen refinement algorithm for this study. After refinement, it is a single pixel width, the intersection is more intuitive and realistic than the first two, and the refinement effect is better.

Figure 4a–c compared to the intersection topology changes, mainly because the refinement process not only depends on the previous iteration results but also depends on the iteration that has been processed pixel distribution. The number of iterations, neighborhood size, and refinement conditions significantly relate to one another. The essence of the three is to refine the multi-pixel width section into a central line. Although the local details, such as intersections, may be different, the overall direction of the section is consistent. There are still isolated and fewer pixels in the refined image, which is caused by noise such as drift or missing trajectory data. For extracting road segments, this part of the pixel is worthless and needs to be filtered.

2.3. Extracting Road Segments from Refined Road Network to Construct Vector Road Network

The ranks of pixels are their coordinates, and pixel coordinates represent the extracted road sections, road intersections, and endpoints. Before extracting the road section, it is necessary to obtain the road section’s end points and intersection points. The idea is to determine the number of other target pixels in the 3 × 3 neighborhood of a target pixel (pixel value is 0). If there are 2, the pixel is the endpoint, and more than 3 are the intersection points.

The road section extraction is divided into two categories. The first is the road section from the endpoint to the endpoint or the intersection. For example, in Figure 4, the number is one part of the road section, and the number is three parts of the intersection. As the refined image is a single pixel width, it can start from each endpoint, in turn, in the 3 × 3 neighborhood to find and store points with gray values, and then start with the stored points and repeat the search. When other intersections or endpoints stop, the road section can be extracted from the endpoint to the endpoint or intersection. The second is the road section between the intersection and intersection points. As shown in Figure 5, the number is two, starting from where each pixel value in the 3 × 3 neighborhood of a specific intersection is 0, similar to extracting the road section from the endpoint. However, when another intersection stops, all road sections between the intersection point and the intersection point can be extracted.

Based on the endpoints and intersections, all the road sections in the refined image are extracted, but the road sections composed of fewer pixels are not necessary. This type of road section is mainly composed of the road sections between the adjacent intersections in the image (which are the same road intersections in reality). These road sections need to be deleted, and then these intersections are merged into one intersection, and the dark part is the intersection, as shown in Figure 6b.

After the test, set to delete the lengths of less than or equal to 4 pixels of the road, most of which are closer to the intersection of the road section between the changes before and after the removal of some of the intersections; Figure 6a,b show two different intersections.

After deleting the short-circuit section, the point must be a natural intersection if an intersection is connected to more than three sections simultaneously. According to this idea, this kind of intersection is obtained, such as in Figure 6a. After this step, some fundamental intersection points are obtained; secondly, if there are other intersections around the intersection, as shown in Figure 6b, after deletion, it is only necessary to start from a particular intersection, search for other intersections within the neighborhood of the intersection, and use these intersections to interpolate a new intersection. The intersection in these road sections is replaced with a new intersection. One intersection can be connected to multiple road sections to obtain a natural intersection after merging, as shown in Figure 6b. The road section is better connected after merging the intersection, and a natural intersection is connected to at least three road sections.

Before constructing the vector road network, the DP (Douglas–Pucker) algorithm is used to compress the road segments. The DP algorithm makes the original curve closer to the natural road and maintains a smooth shape, but an appropriate threshold must be selected [24,25,26,27,28,29]. The threshold selected in this study is a 0.8-pixel width, which represents the accurate distance of 4 m. The DP algorithm compresses the extracted road section. Since the endpoint of the compressed road section must be the endpoint or the real intersection point, it is only necessary to generate vector road sections for each road section in turn. GPS trajectory data contain latitude and longitude coordinates, which are fitted after data processing and combined with the road network coordinates of the shp file exported by OSM, by calling the GDAL library, reflected in ArcGIS.

3. Results

The script for generating trajectory points is written in the Python language. The input parameters are the grid coordinate points of the research area, and the returned parameters are trajectory points p. The trajectory points are generated in ArcGIS software, and then the road network of the study area is obtained by rasterizing the trajectory and vectorizing the raster data. The study area is Harbin City in Northeast China, with a total area of 10,249 km². The road network of the main urban area of Harbin was obtained, as shown in Figure 7.

To evaluate the effect of the extracted road network, this study first overlays the extracted road network with the OSM road. As shown in Figure 8, it can be seen that the OSM road network can reflect the road network structure of the main roads in the main urban area of Harbin, but there are severe deficiencies in the secondary roads and branches. The road network extracted in this paper and the existing roads of OSM also added a more detailed road network structure. To further evaluate the correspondence between the extracted road network and the existing road network, the extracted road network is compared with the map of the network (Figure 8). On the whole, the extracted road network can better reflect the topological structure of the urban road network, and the roads at all levels have a high degree of coincidence. The network map is used as a reference to manually vectorize a small part of the road network as the evaluation area, and the buffer method proposed in [30,31] is used to evaluate the road network accuracy of the evaluation area. A specific range of buffers is made for the extracted road network. The extraction accuracy is the ratio of the actual extracted road network length falling into the buffer zone to the total length of the existing road network. As the road network extracted intuitively has a high degree of agreement with the network, a smaller buffer with a radius of 5 m is set, and a road network extraction accuracy of 91.39% is obtained.

4. Discussion

In this study, 3,515,387 taxi trajectory data from 202,205,011 to 202,205,018 in a 5 km × 5 km area of Harbin were used for experiments. The minimum grid cell is 5 m × 5 m, the traditional grid method needs 22,749 ms, and the hierarchical rasterization only needs 21,317 ms. Compared with the two, the speed of hierarchical rasterization is improved. The road network of the selected area after hierarchical rasterization, expansion, refinement, and vectorization after denoising is shown in Figure 9.

Compared with Figure 9b, the road centerline in Figure 9a is obvious. The adjacent lanes are merged, and the road sections are disconnected, which is more conducive to extracting the road centerline, but the road centerline is a non-single pixel width. Figure 9c shows the region’s refined and denoised road network image. The road width is a single pixel width consistent with the road network topology information in Figure 9b. Figure 9d is the vector road network extracted based on Figure 6a. The road center lines in the two figures are the same, which shows that the generation process of the vector road network is correct. To illustrate the effectiveness of the generated road network, this study selects two areas in the vector road network to compare with the corresponding network map road network. As shown in Figure 10, the road sections in the vector road network are close to the road sections in the entire map road network. The topological structure characteristics of the road sections are the same. However, there are some missing road sections, mainly because the trajectory data do not cover these areas or the number of trajectory points in the area is small. Then, the latitude and longitude coordinates of the extracted real intersection point are compared with the actual road intersection point coordinates. Points A, B, and C of area 2 in Figure 10 are selected. The comparison results are shown in

Table 3. Comparison of intersection coordinates in region 2.

ID	Coordinates in (a) (°)		Coordinates in (b) (°)		Gap/m
A	126.573473	45.747516	126.573195	45.747477	31
B	126.558710	45.749762	126.558571	45.749898	22
C	126.584374	45.743233	126.584262	45.743435	26

. By comparison, the coordinate difference is about 30 m, and this kind of error is mainly caused by turning the intersection into a point in extracting the intersection, which further proves that this study is feasible to extract the road network from the taxi trajectory data.

5. Conclusions

The vector road network data in the existing land spatial planning are challenging to obtain but are particularly needed in project planning, especially at the city level. The trajectory generation and road network extraction method based on map API proposed in this study has a good extraction effect, effectively solving the road network acquisition problem in spatial planning. Additionally, it provides a new idea for fine-scale road network extraction research. The method proposed in this study has the following advantages: (1) The trajectory generation method based on map API provides a new data source for the study of road network extraction; (2) This method is not limited to urban-level trajectory generation and road network extraction, but also applies to trajectory generation and road network extraction in smaller areas. For small areas, it is necessary to change the driving path service of the map API to the walking path service; (3) The existing Internet map API data are open-source sound, have low development difficulty, are convenient for non-computer professionals to learn quickly, and are conducive to the rapid development of the project and research. In addition, aiming at the problem of constructing a vector road network by using trajectory data, this paper proposes to use hierarchical rasterization and an improved Zhang–Suen refinement algorithm to realize fast extraction of the road network. The proposed road network generation algorithm only calculates the road’s center line and does not obtain detailed information such as the number of lanes. In future work, refined attributes such as the lane number and lane direction can be determined according to the distribution characteristics of GPS trajectory.