International Journal of
Geo-Information Article
Incremental Road Network Generation Based on Vehicle Trajectories Zhongyi Ni 1 , Lijun Xie 1, *, Tian Xie 2 , Binhua Shi 1 and Yao Zheng 1 1 2
*
School of Aeronautics and Astronautics, Zhejiang University, Hangzhou 310027, China;
[email protected] (Z.N.);
[email protected] (B.S.);
[email protected] (Y.Z.) Zhejiang Lab, Hangzhou 311121, China;
[email protected] Correspondence:
[email protected]; Tel.: +86-131-8500-1081
Received: 11 August 2018; Accepted: 21 September 2018; Published: 23 September 2018
Abstract: Nowadays, most vehicles are equipped with positioning devices such as GPS which can generate a tremendous amount of trajectory data and upload them to the server in real time. The trajectory data can reveal the shape and evolution of the road network and therefore has an important value for road planning, vehicle navigation, traffic analysis, and so on. In this paper, a road network generation method is proposed based on the incremental learning of vehicle trajectories. Firstly, the input vehicle trajectory data are cleaned by a preprocess module. Then, the original scattered positions are clustered and mapped to the representation points which stand for the feature points of the real roads. After that, the corresponding representation points are connected based on the original connection information of the trajectories. Finally, all representation points are connected by a Delaunay triangulation network and the real road segments are found by a shortest path searching approach between the connected representation point pairs. Experiments show that this method can build the road network from scratch and refine it with the input data continuously. Both the accuracy and timeliness of the extracted road network can continuously be improved with the growth of real-time trajectory data. Keywords: road network; incremental learning; vehicle trajectories
1. Introduction Street maps and transportation networks are the bases of building smart cities. High accurate road network maps have great social and application values. Currently, road network generating and updating algorithms can be mainly divided into three categories. (i) Field measurement based on the professional GPS equipment and surface measurement technology [1]: This method relies on the professional road measurement vehicles and data collection personnel. It suffers the disadvantages of long work cycle, unstable measurement accuracy, high cost, expensive to maintain, etc. With the development of the satellite technology, its application range becomes even smaller. (ii) Extracting road network map from remote sensing image based on the image processing technologies [2–4]: This method relies on remote sensing. However, high-definition remote sensing maps have low real-time performance and high purchase costs. They are limited by the image processing technology and are difficult to automate. Therefore, the extraction efficiency is relatively low. (iii) Building the road network with Volunteered Geographic Information (VGI) [5]: This method relies on VGI, which is the harnessing of tools to create, assemble, and disseminate geographic data provided voluntarily by individuals. Thus, the quality of the updated map depends on the skill level of the volunteer and the accuracy of the data.
ISPRS Int. J. Geo-Inf. 2018, 7, 382; doi:10.3390/ijgi7100382
www.mdpi.com/journal/ijgi
ISPRS Int. J. Geo-Inf. 2018, 7, 382
2 of 19
In recent years, with the rapid development of global positioning systems (GPS), radio frequency identification technologies (RFID), and sensor technologies, it has become easier to collect the location information of moving objects. Nowadays, more and more cars have installed GPS devices to record the vehicle’s trajectory. These vehicles can generate tens of thousands of vehicle trajectory data every day. These vehicle trajectory data not only contain the laws of the vehicle’s movements and traffic congestion, but also reveal the shape of the road network and the rules of the road network’s evolution over time. Therefore, more and more researchers have begun to use vehicle trajectory data to generate and update the road network information. The existing methods for constructing a road network using vehicle trajectory data can be roughly divided into three categories. (i) Point clustering assumes that the input is a set of points and clusters the points together through different clustering methods to obtain a road network map [6–11]. Representative algorithms in this class include the following. Li et al. [8] proposed the use of spatial-linear clusters to infer road segments from GPS trajectories. Their algorithm can detect missing road and checking the correctness of existing road network through inferring road segments. Edelkamp et al. [9] clustered high precision DGPS trajectories to construct road network, and the center of each cluster is regarded as the lane center line. In [10], GPS points are converted into binary image by morphological operations. Then, the skeleton is extracted to construct road network. Chen et al. [11] proposed a map interface algorithm with accuracy guarantees based on detecting seed elements and connecting them subsequently. (ii) Incremental track insertion uses the idea of map matching to gradually insert the trajectory into the initial map to construct a road network map [1,12–14]. Representative algorithms in this class include the following. Zhang et al. [12] combined the K-Means clustering with the Gaussian model to extract the centerline of the road and continuously refine existing road network. Bruntrup et al. [13] proposed a spatial-clustering based algorithm that allows incrementally generating a road network, but the algorithm requires high quality (sampling rate and positional accuracy) tracking data. Cao and Krumm [1] applied a custom clustering algorithm to group similar input trajectories together and then build up the road network incrementally. Ahmed et al. [14] proposed an incremental algorithm for the road network construction that matching of trajectories and map is achieved by Fréchet distance. Although this algorithm guarantees the local quality of the road network, it does not solve the basic connectivity problem. (iii) Intersection linking determines the intersection through the motion characteristics (speed and direction) or point density of the vehicle, and then connects the intersections by interpolation [15,16]. Representative algorithms in this class include the following. Fathi and Krumm [15] introduced an intersection detector trained on ground truth data from an existing map. Firstly, they find the intersections using a classifier learned over the shape descriptors. Then, they connected the intersections with geometrically accurate road segments. Finally, they used the iterative closest point algorithm to optimize the position of each intersection. Karagiorgou et al. [16] proposed the Trace Bundle algorithm, which realizes the classification of the trajectory by intersection turning model, and using trajectory clustering to realize road network extraction. In this work, an algorithm for incrementally learning vehicle trajectory data and generating a road network is proposed. The algorithm does not require the existing road network as the basis. It incrementally generates a road network by learning the position and timing information from the input vehicle trajectory. The road network graph has a high timeliness and can be continuously updated when the input trajectory changes. The outline of the paper is depicted as follows. Section 1 describes a detailed analysis of the characteristics of the vehicle trajectory data and explains the advantages of using the vehicle trajectory data to extract the road network. Section 2 depicts the detailed specific implementation flow of the road network extraction method based on incremental learning. Section 3 shows the experimental results and comparative analysis. Finally, Section 4 discusses conclusions and future work.
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW
3 of 19
2. Analysis of Characteristics of Vehicle Trajectory Data A trajectory a polyline in a multidimensional space formed by a series of sampling ISPRS Int. is J. Geo-Inf. 2018, 7, 382 3 of 19points that contain attributes such as geographical location and time, which are used to represent the positional change of an object over a period. Vehicle trajectory refers to the trajectory of a set of sampling points 2. Analysis of Characteristics of Vehicle Trajectory Data obtained by the vehicle-mounted GPS device in the journey. The sampling points of a vehicle A trajectory is a polyline in a multidimensional space formed by a series of sampling points that trajectory generally contain attributes such as and position, time, and direction. High-quality contain attributes such as geographical location time, which are speed, used to represent the positional change ofdata an object a period. social Vehicleand trajectory refers to the trajectory of a set of sampling vehicle trajectory haveover important application values in solving social issues such as points obtained by the vehicle-mounted GPS device in the journey. The sampling points of a vehicle traffic congestion, traffic services improving, road environment monitoring, and energy shortages trajectory generally contain attributes such as position, time, speed, and direction. High-quality alleviatingvehicle [17]. In this paper, theimportant road network isapplication extractedvalues frominthe vehicle data mainly trajectory data have social and solving socialtrajectory issues such as based on the following characteristics of the vehicle trajectory:monitoring, and energy shortages traffic congestion, traffic services improving, road environment
alleviating [17]. In this paper, the road network is extracted from the vehicle trajectory data mainly
The trajectory data express the information of road structure. The movement of a typical vehicle based on the following characteristics of the vehicle trajectory: is always limited to the existing road network (the vehicle cannot move freely on the plane), so • The trajectory data express the information of road structure. The movement of a typical vehicle the vehicle trajectory data represent richnetwork road structure In Figure 1, many vehicle is always limited to the existing road (the vehicleinformation. cannot move freely on the plane), trajectories superimposed, roughly the structure road network so are the vehicle trajectory data representdelineating rich road structure information.of Inthe Figure 1, many vehiclein the area. trajectories are superimposed, roughly delineating the structure of the road network in the area. the rapid The real-time vehicle trajectory data express the dynamic changes of road status. With • The real-time vehicle trajectory data express the dynamic changes of road status. With the rapid development of the city, urban roads are changing from time to time because of road development of the city, urban roads are changing from time to time because of road construction, construction, road maintenance, traffic high real-time performance is road maintenance, traffic control, etc. control, Therefore,etc. highTherefore, real-time performance is required for required for the road network extraction The vehicle by the road network extraction algorithm. algorithm. The vehicle trajectories are trajectories determined byare the determined road status which is always changing along with the road transformation. Therefore, it is a unique the road status which is always changing along with the road transformation. Therefore, it is a advantage to utilize the vehicle trajectory data to update the road network. unique advantage to utilize the vehicle trajectory data to update the road network.
Figure 1. Vehicle trajectoriesduring during aacertain period of a certain area. Figure 1. Vehicle trajectories certain period of a certain area.
3. Extraction of Road Network by Incremental Learning Method
3. Extraction ofFigure Road Network by Incremental Learning Method 2 depicts the algorithm flow of incrementally learning from the vehicle trajectory data building a road network. The algorithm has an input port to continuously receive trajectory data. Figureand 2 depicts the algorithm flow of incrementally learning from the vehicle trajectory data Whenever a vehicle trajectory is obtained, firstly it will be pre-processed to ensure the correctness of and building road network. Thethe algorithm has an input toinformation continuously receive trajectory data. the atrajectory data. Then, position information and port timing is learned from the trajectory online. Finally, the roadfirstly network generated or updated incrementally on Whenever input a vehicle trajectory is obtained, it is will be pre-processed to ensurebased the correctness of the information. the trajectory data. Then, the position information and timing information is learned from the input trajectory online. Finally, the road network is generated or updated incrementally based on the information.
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW
4 of 19
ISPRS Int. J. Geo-Inf. 2018, 7, 382
4 of 19
Representative point Truck trajectory
extraction
Taxi trajectory Private car trajectory Bus trajectory
Incremental generation of
pretreatment
road network Connecting segment extraction Figure 2. Road network extraction algorithm flow. Figure 2. Road network extraction algorithm flow.
In the following section, the four steps of the proposed algorithm, including trajectory data In the following section,point the four steps ofconnecting the proposed algorithm, including trajectory data preprocessing, representative extraction, segment extraction and road network preprocessing, representative point in extraction, incremental generation, are described detail. connecting segment extraction and road network incremental generation, are described in detail. 3.1. Preprocessing of Vehicle Trajectory Data 3.1. Preprocessing of Vehicle Trajectory Data The GPS point herein refersrefers to theto coordinate point of point the vehicle position Definition Definition1.1.GPS GPSpoint. point. The GPS point herein the coordinate of the vehiclemeasured position bymeasured the vehicle-mounted GPS device during the running of the vehicle. It is expressed as p = {x, y, t}, where by the vehicle-mounted GPS device during the running of the vehicle. It is expressed asxp and y, y, respectively, to the latitude and longitude the vehicle which is measured by the = {x, t}, where xcorrespond and y, respectively, correspond to theoflatitude andlocation, longitude of the vehicle location, positioning device, and tby indicates the time when the vehicle in this position. which is measured the positioning device, and t is indicates the time when the vehicle is in this position. Due to the limitations of the positioning accuracy and signal strength of the GPS device, there are usually a large number ofoferrors and information missing in thestrength original of vehicle trajectory Due to the limitations the positioning accuracy and signal the GPS device,data. there Therefore, it aislarge necessary to of preprocess input vehicle trajectory data to repair remove data. the are usually number errors andthe information missing in the original vehicleor trajectory abnormal data. pi , pi+1 be adjacentthe GPS points in a trajectory, pi = ti ), pi+1 the = ( xi , yori , remove Therefore, it isLet necessary to two preprocess input vehicle trajectorywhere data to repair , y , t . There are three typical cases on how the errors are introduced: ( xabnormal ) data. Let 𝑝 , 𝑝 be two adjacent GPS points in a trajectory, where 𝑝 = (𝑥 , 𝑦 , 𝑡 ), 𝑝 i +1 i +1 i +1 𝑖 𝑖+1 𝑖 𝑖 𝑖 𝑖 𝑖+1 = , 𝑦𝑖+1 , 𝑡a𝑖+1 ). There threeon typical cases the on how theoferrors are introduced: 1.(𝑥𝑖+1When vehicle is are driving the road, signal the GPS positioning equipment may be or interrupted because the occlusion of of trees, and other objects 1. disturbed When a vehicle is driving on theofroad, the signal the high GPS buildings positioning equipment mayon be both sides of the road, or vehicles entering tunnels, underground parking areas, etc., resulting disturbed or interrupted because of the occlusion of trees, high buildings and other objects on inboth the sides interruption of theorvehicle trajectory. 3 depicts an example of such missing points. of the road, vehicles enteringFigure tunnels, underground parking areas, etc., resulting Ifinp3 and p4 are directly connected, an erroneous trajectory will be formed. This work judges the interruption of the vehicle trajectory. Figure 3 depicts an example of such missing points. whether there in trajectory according to the time between adjacent If p3 and p4 are missing directlypoints connected, an erroneous trajectory willinterval be formed. Thistwo work judges GPS pointsthere (|ti+1are − timissing and discards the trajectory of ato serious loss interval [18]. | > tmax ),points whether in trajectory according the time between two 2. The GPS device signal(|𝑡 interference cause positioning data toloss deviate adjacent GPS points 𝑡𝑚𝑎𝑥additionally ), and discards theGPS trajectory of a serious [18]. from 𝑖+1 − 𝑡𝑖 | > will vehiclesignal driving routes and form noise points, asGPS depicted in Figure points 2. the Theactual GPS device interference will additionally cause positioning data4.toNoise deviate from will affect the shape of theroutes trajectories andnoise make it failastodepicted match the actual4.route. noise the actual vehicle driving and form points, in Figure NoiseThe points will points in the trajectory can be found and itremoved according to the average speed [18],in affect the shape of the trajectories and make fail to match the actual route. The noise points which can be calculated by the distance and time interval between two adjacent GPS points the trajectory can be found and removed according to the average speed [18], which can be r calculated by the distance and time interval between two adjacent GPS points ( vavg = xi2+1 − xi2 + y2i+1 − y2i /|ti+1 − ti |). (vavg = 2 2 √(𝑥𝑖+1 − 𝑥𝑖2 ) + (𝑦𝑖+1 − 𝑦𝑖2 )⁄|𝑡𝑖+1 − 𝑡𝑖 |). 3.3. When GPS devices may maintain working status. Therefore, there maymay be Whenthe thevehicle vehicleisisstopped, stopped, GPS devices may maintain working status. Therefore, there many same (similar) positioning points in the trajectory data during a long period of time. be many same (similar) positioning points in the trajectory data during a long period of time. Such Suchkind kindofofpoints pointsare arecalled calledstationary stationarypoints points[18], [18],asasdepicted depictedininFigure Figure5.5.InInthe theanalysis analysisofof the road network structure, the stationary points in the vehicle trajectory will generate the road network structure, the stationary points in the vehicle trajectory will generateaalarge large amount points is that the positions ofof amountof ofdata dataredundancy. redundancy.The Themain mainfeature featureofofthe thestationary stationary points is that the positions r 2 2 2 (𝑦 2 − 2 𝑦 2 )
