2020 5th IEEE International Conference on Big Data Analytics

Traffic Intersection Detection Using Floating Car Data

Rong Hu Ye Xia
Fujian Provincial Key Laboratory of Automotive School of Ecological Environment and Urban
Electronics and Electric Drive, Construction, Fujian University of Technology
Fujian University of Technology Fuzhou, China
Fuzhou, China e-mail: [email protected]
e-mail: [email protected]

Chih-Yu Hsu Hanlin Chen, Weihui Xu

School of Information Science and Engineering Fujian Provincial Key Laboratory of Automotive
Fujian University of Technology Electronics and Electric Drive,
Fujian Provincial Key Laboratory of Big Data Mining Fujian University of Technology
and Applications National Demonstration Center Fuzhou, China
Fuzhou, China e-mail: [email protected],[email protected]
e-mail: Corresponding Author: [email protected]

Abstract—The autonomous vehicle navigation system and the considerable amount of GPS trajectories named Floating Car
intelligent traffic system need real-time routing planning in Data (FCD).
traffic engineering and management. Urban road intersections Traffic regulation information plays a significant role in
play bottlenecks of urban roads, contributing a great deal to the traffic control system for the urban environment. Traffic
the loss of travel time due to traffic interference, management Navigation System can guide vehicles to run on the optimal
and control. The detection of the traffic intersection is path by avoiding congested roads and increasing the capacity
important for obtaining real-time road regulation information. of the existing road network. However, traffic intersections
This paper proposes a method to detect the traffic intersection play an important role in the control of traffic congestion.
based on floating car data that is typically timestamped geo-
Urban road intersections play an important role in deciding
localization, direction and speed data directly collected by
moving vehicles. The angle of the direction difference can be
the total travel time and the overall travel efficiency because
used as the features for detecting the traffic intersection. The they are the meeting points of pedestrians, bicycles and
identification of the traffic intersections algorithm is developed vehicle flow, and junctions where road users change their
based on the data density clustering algorithm. The floating directions. They are also the bottlenecks of urban roads,
car data is collected by about 15,000 cars using an onboard contributing a great deal to the loss of travel time due to
with the Global Positioning System receiver and sensors that traffic interference, management and control [3, 4]. By using
providing positions (latitude, longitude) instantaneous speed, traffic cameras, streams of trajectory points [5] can be
drive direction and time etc. The assessment results show the collected from the vehicles on the road. The outliers among
approach has excellent performance with high accuracy for the trajectory streams are possible to find those vehicles
traffic intersection detection. whose trajectories significantly deviate from the normal
driving behaviors [6, 7, 8, 9] that increase the traffic
Keywords- Traffic intersection detection; Mobile Trajectory congestion. There are some types of traffic intersections like
of Vehicle; Floating car data; Intelligent traffic system four-way, T-junction, Y-intersection, traffic circle, fork, and
Turning lanes. There are four situations of turning
I. INTRODUCTION trajectories at the four-way traffic intersection like going
Recently researches are focusing on developing the straight, left turn, right turn and, U-turn.
autonomous vehicle which can accept these noisy inputs and Li et al. [7] first proposed the stay point detection
still navigate and localize itself correctly on the road. [1].ġ algorithm. This algorithm first checks if the distance between
Real-time path planning is one of the key technologies in an anchor point and its successors is in a trajectory larger
autonomous driving. The floating cars are vehicles like than a given threshold (e.g., 100 m). This idea also initiates
this work to propose an algorithm for the detection of traffic
Global Position System (GPS)-equipped taxi, bus, interurban
intersections based on FCD. The algorithm can locate traffic
coach. They are an effective way to collect real-time traffic
intersections on the map of Fuzhou city. The algorithm
information data in the city road network. The traffic and
includes a density-based clustering algorithm calculating the
transportation sector is experiencing a massive increase in
angle difference of the trajectory data derived by analyzing
the amount of traffic data (e.g., vehicles’ travel trajectory
FCD. The trajectory data collected from the cars drive with
data) collected. [2]. The real-time data collected by GPS
the change of directions that are not always at the
devices on the moving vehicles are real-time, collecting a

978-1-7281-4111-4/20/$31.00 ©2020 IEEE 116

Authorized licensed use limited to: University of Wisconsin. Downloaded on April 23,2025 at 02:49:36 UTC from IEEE Xplore. Restrictions apply.
intersections. These kinds of trajectory data belong to A. Visulization of the Trajectory Data
outliers of the suspect traffic intersections. Density Based The Trajectory of the car is shown in Fig. 1. The car is
Spatial Clustering of Applications with Noise (DBSCAN) is driving across two intersections. The records obtained by
a density-based clustering algorithm to detect outliers of the filtering the FCD with ID=55326 is in Table II. The column
suspect traffic intersections [10]. Lon is the values of the values of Longitude. The column
II. FLOATING CAR DATA LAT is the values of Latitude. The column SPEED is the
speed value of the car that is measured by km/hr.The column
The FCD dataset was collected in 2018/5/5 in Fuzhou DIR is the value of the car head direction that is measured by
city. There is a sample data file contains ten examples of degree unit that is angle measurement in general.
records describing information, as follows in Table I.

TABLE I. FCD DATASET

ŊŅġ ōŐŏġ ōłŕġ ŔőņņŅġġ ŕŊŎņġ ŅŊœġġ

2018/5/5
73042 119.2964 26.08945 0 269
0:17:40
2018/5/5
65735 119.3384 26.0666 21 225
0:17:40
2018/5/5
35982 119.353 26.09887 0 175
0:17:40
2018/5/5
53427 119.2925 26.0697 21 270
0:17:40
2018/5/5
73146 119.3137 26.08554 47 152
0:17:40
2018/5/5
63121 119.3012 26.1052 23 180
0:17:40
2018/5/5
55326 119.2881 26.07553 50 274
0:17:40
2018/5/5
62312 119.3116 26.10138 19 164
0:17:40
2018/5/5 Figure 1. The trajectory of the car with the ID=55326
66637 119.3272 26.07279 0 282
0:17:40
2018/5/5 Fig. 2 shows three situations of turning trajectories at one
66122 119.3144 26.07984 40 173
0:17:40
traffic intersection.
Where the DIR represents the direction of one specific
taxi car ID=55326, there are only 10 records of trajectory
data of the vehicle is listed as examples in Table II. After
sorting data records by timestamp, the two adjacent records
represent two location points of the vehicle located nearby.
III. DATA PROCESSING AND ANALYSIS
The directions of trajectories of cars passing intersections
have some features that can be used to develop a recognition
algorithm of intersections. When cars are turning right or left
to pass through the intersections, there are only a few data
obtained because the sampling rate of FCD is one record for
every 20 seconds. The data of trajectories show the
differences of directions are candidates to differentiate
samples that can identify the road intersections to improve
the computation efficiency. The following subsections
describe types of differences of directions for developing the
DBSCAN-based algorithm of road intersection identification.
The car sometimes goes into the gas station to refuel, and
there will be a change in the track point in its trajectory. The Figure 2. Three situations of turning trajectories
characteristic of this kind of path is that the driving direction
will change frequently, the speed will decrease until it is zero, There are two records at times 1:26:28 and 1:26:48
or the rate will increase from zero, the route will change, the obtained, and the period of sending the data is 20 seconds.
speed will increase until it is stable, and the direction will be The direction of the car is forward to the east and the
steady. Fig. 1 shows an example of trajectory for one car directions (DIR) are 76 and 92. As shown in the Fig. 2, the
with the ID=55326. direction is 90 if it’s direction is forward to the right East.

117

Authorized licensed use limited to: University of Wisconsin. Downloaded on April 23,2025 at 02:49:36 UTC from IEEE Xplore. Restrictions apply.
 TABLE II. RECORDS FILTERED BY ID=55326 TABLE III. SOME LOCATIONS OF THE NODES NEAR THE
INTERACTIONS

No. LON LAT SPEED TIME DIR ŏŰįįġ ōŐŏġġ ōłŕġġ

4 119.3116 26.07954
2018/5/4
1 119.3078 26.07967 46 76
1:26:28 5 119.314 26.0796
2018/5/4 6 119.3141 26.07962
2 119.3107 26.0795 41 92
1:26:48
7 119.3144 26.07968
2018/5/4
3 119.3111 26.07953 0 106 8 119.3144 26.08179
1:27:08
9 119.3143 26.08368
2018/5/4
4 119.3116 26.07954 35 86
1:27:28
2018/5/4 The six points have one candidate point 7 to represent the
5 119.314 26.0796 15 88
1:27:48
intersection because point 7 is a turning point. The values of
2018/5/4 Longitude (Lon) in No. 8 and 9 vary not as much as the
6 119.3141 26.07962 0 0
1:28:08 values of Latitude(LAT) because the car is driving from the
7 119.3144 26.07968 22
2018/5/4
66
South to the North.
1:28:28 How can we find the candidate point 7? The rules can be
2018/5/4 found in Table IV. The values of the index DD at No. 4, 5
8 119.3144 26.08179 50 358
1:28:48 are less than 25 and The values of the index DD at 6 and 7
2018/5/4 are larger than 25. Three thresholds for three parameters,
9 119.3143 26.08368 20 10 Difference of Time (DT), the direction change (DD),
1:29:08
2018/5/4
difference of location (DL). According to the previous
10 119.3166 26.084 46 88 analysis, when the vehicle turns at the intersection, there are
1:29:28
generally 1 or 2 points where the direction changes, the
change value is usually above 25 degrees, the time is
typically 20-40 seconds, and the distance is no more than
B. Computation of Indexes 200 meters.
There are three indexes (DT, DD, DL) are introduced in
the subsection. The difference of time between two time- TABLE IV. COMPUTATION OF INDEXS
stamps at ith and i+1th of the car data records denotes DT(i),
and it is formulated as in equation (1). TIME DIR DT DD DL

1 76
20 16 290
DT i =TIME i+1 -TIME i ġ (1) 2 92
3 106 20 14 40
The difference of direction between two time-stamps at ith 20 20 50
4 86
and i+1th of the car data records denotes DD(i), and it is
5 88 20 2 239
formulated as in equation (2).
6 0 20 88 10
DD i =DIR i+1 -DIR i ġ (2) 20 66 30
7 66
The location difference between two time-stamps at ith 8 358 20 292 234
and i+1th of the car data records denotes DL(i) and it is 9 10 20 348 210
formulated as in equation (3). 20 78 232
10 88
DL i =Location i+1 -Location i ġ (3)

where the location is related to Degrees Minutes Seconds C. Intersection Detection Algorithm
(DMS) that are converted from GPS Coordinates, Latitude, The car is driving across two intersections with the
and Longitude. The distance can be calculated by using the change of directions, as shown in Fig. 1. The Trajectory of
‘haversine’ formula. the car shown in Fig. 1 has the state of data at any moment.
According to Fig. 1, the values of DIR are approximating The speed is reduced to zero when the car is driving across
90 if the car drive from the West to the East and the values of two intersections and the direction is changed. The change of
DIR are approximating 0 if the car drive form the South to direction DD(i) is shown in Table IV. For turning left, the
the North. In Table III, the values of Latitude(LAT) vary not vale of DD is changed from East to North in Fig. 1. Fig. 3 is
as much as the values of Longitude (Lon) because the car is the flowchart of the intersection detection algorithm.
driving from the West to the East like the NO 4,5, 6 and 7. The first rule is any value of DD(i) larger than 25. The
second rule is the value of DD(i) should be approximated 90

118

Authorized licensed use limited to: University of Wisconsin. Downloaded on April 23,2025 at 02:49:36 UTC from IEEE Xplore. Restrictions apply.
and the value DT should be between 20 and 60, and the The cars drive with the change of directions that are not
value DL should be between 50 and 200. always at the intersections, as shown in Table I. These kinds
of trajectory data belong to outliers of the suspect traffic
intersections. When the changes of directions used to
identify the traffic intersections, it is necessary to filter out
the trajectory data that does not belong to the intersection.
The DBSCAN cluster algorithm is a density-based clustering
algorithm to detect outliers of the suspect traffic intersections.
It is focused on finding neighbors by density (MinPts) on an
‘n-dimensional sphere’ with radius ɛ (epsilon). A cluster can
be defined as the maximal set of ‘density connected points’
in the feature space, as shown in Figure 5a. The DBSCAN
cluster algorithm has two parameters, ɛ and the minimum
number of points (MinPts), as shown in Figure 5b. These
two parameters are chosen ɛ=35 and MinPts=7 by
experience.

Border 
Point


Core
Point

(a) (b)

Figure 5. The DBSCAN cluster algorithm has (a) border point and core
point (b) parameters radius H and MinPts=7.
Figure 3. The flowchart of the intersection detection algorithm
where H represents the radius and the value range.
DBSCAN captures the clusters are dense groups of points.
Figure 6 shows one day for all the car trajectory points that at
the intersection after filtering by the flowchart in Figure 3.
All car trajectory points of one day in Fuzhou City are
matching on the road map of Fuzhou, China in the latitude
and longitude range (119.29958,26.09649)-
(119.32019,26.08425) and they are shown in Figure 6.

Figure 4. The detected points are candidates of the intersections.

Figure 4 shows all the car trajectory points that at the

intersection after filtering by The flowchart of the
Figure 6. The red points are centroid points of clusters obtained by the
intersection detection algorithm for one day. The detected DBSCAN cluster algorithm without removing the candidate points.
points are candidates of the intersections. The DBSCAN
cluster algorithm

119

Authorized licensed use limited to: University of Wisconsin. Downloaded on April 23,2025 at 02:49:36 UTC from IEEE Xplore. Restrictions apply.
 ACKNOWLEDGMENT

This work was supported by Fujian Provincial

Department of Science and Technology (Grant No.
2017J01729)

REFERENCES
[1] Massachusetts Institute of Technology. "Bringing human-like
reasoning to driverless car navigation: Autonomous control system
'learns' to use simple maps and image data to navigate new, complex
routes."ScienceDaily.www.sciencedaily.com/releases/2019/05/19052
3130212.htm (accessed January 1, 2020).
[2] Guo, F., Zhang, D., Dong, Y. et al. Urban link travel speed dataset
from a megacity road network. Sci Data 6, 61 (2019)
Figure 7. The red points represent the location of intersections. doi:10.1038/s41597-019-0060-3
[3] Deqi Chen, Xuedong Yan, Feng Liu, Xiaobing Liu, Liwei Wang, and
Jiechao Zhang, "Evaluating and Diagnosing Road Intersection
IV. RESLUTS AND DISCUSSIONS Operation Performance Using Floating Car Data," Sensors (Basel).
2019 May; 19(10): 2256. Published online 2019 May 15. doi:
According to the previous analysis, some turning points 10.3390/s19102256
of trajectory exist at the intersection. It is reasonable to [4] Heidemann D. Queue length and delay distributions at traffic signals.
detect an intersection by the turning points from the Transp. Res. Part B, Methodol. 1994;28:377–389. doi: 10.1016/0191-
trajectory data. By calculating the difference in the direction 2615(94)90036-1.
of two continuous track points (direction change DD). We [5] Y. Ge, H. Xiong, Z.-h. Zhou, H. Ozdemir, J. Yu and K. C. Lee, "Top-
Eye: top-k evolving trajectory outlier detection," in CIKM '10
can judge those turning points by the trajectory of a car. Proceedings of the 19th ACM international conference on
Through the previous analysis, we also know that there are 4 Information and knowledge management, 2010.
cases in this direction change. We also have to select the [6] E. Leal, and L. Gruenwald. “Research Issues of Outlier Detection in
track points that belong to the intersection from the trajectory Trajectory Streams Using GPUs.” In: SIGKDD Explore. News. 20.2
data, and filter out those are not part of the intersection but (Dec. 2018), pp. 13–20. ISSN: 1931-0145. doi:
also have the characteristics of direction change. Figure 6 10.1145/3299986.3299989.url:
http://doi.acm.org/10.1145/3299986.3299989
shows an example of a display of the car trajectories points
on the map after non-intersection turning trajectory points. In [7] J. Lee, J. Han, and X. Li. “Trajectory Outlier Detection: A Partition-
and-Detect Framework.” In: 2008 IEEE 24th International
Fig. 6, the point density of the intersection is significantly Conference on Data Engineering. Apr. 2008, pp. 140–149. doi:
higher than the density of other places. Only the red points 10.1109/ICDE.2008.4497422.
are obtained by removing the other candidate points that are [8] Y. Ge, H. Xiong, Z.-h. Zhou, H. Ozdemir, J. Yu, and K. C. Lee.
not core points. The red points represent the location of “Top-Eye: Top-k Evolving Trajectory Outlier Detection.” In:
intersections as shown in Fig. 7. Proceedings of the 19th ACM International Conference on
Information and Knowledge Management. CIKM ̓10. Toronto, ON,
V. CONCLUSIONS Canada: ACM, 2010, pp. 1733–1736. ISBN: 978-1-4503-0099-5. doi:
10.1145/1871437.1871716.url:
This paper proposes a method to detect traffic http://doi.acm.org/10.1145/1871437.1871716
intersections based on floating car data (FCD). Features such [9] Gohar, Usman. (2019). Scalable Techniques for Trajectory Outlier
as angle difference, time difference and location difference Detection. Retrieved from the University of Minnesota Digital
are used to recognize the traffic regulations in intersections. Conservancy, http://hdl.handle.net/11299/206183.
The identification of the traffic intersections algorithm use [10] Martin Ester, Hans-Peter Kriegel, Jörg Sander and Xiaowei Xu, "A
Density-Based Algorithm for Discovering Clusters in Large Spatial
the DBSCAN algorithm to find clusters. The results show a Databases with Noise, " Proceedings of the Second International
map in Figure 7. with red points representing the location of Conference on Knowledge Discovery and Data MiningAugust 1996
intersections. Pages 226–231.

120

Authorized licensed use limited to: University of Wisconsin. Downloaded on April 23,2025 at 02:49:36 UTC from IEEE Xplore. Restrictions apply.