Content uploaded by Zhengbing He
Author content
All content in this area was uploaded by Zhengbing He on May 09, 2019
Content may be subject to copyright.
Research based on high-fidelity NGSIM vehicle trajectory datasets:
A review
Zhengbing He1
Abstract
Next Generation Simulation (NGSIM) program published four high-fidelity trajectory datasets
more than ten years ago. Recognizing the great influence of the datasets on transporta-
tion research, this paper classifies and reviews the research based on the NGSIM trajectory
datasets. Due to the wide existence of relevant literature, only the papers published in the
leading journal series of Transportation Research are considered. Those papers are then
classified into six subjects, and it is found that the data are mainly employed in five ways.
To shed light on the future data usage and collection, limitations of the NGSIM datasets are
pointed out, and outlooks of future data collection are presented.
Keywords: Next Generation Simulation, vehicle trajectory, traffic data, data usage, traffic
flow modeling, transportation research
1Dr. Zhengbing He is always interested in solving various transportation problems by combining
empirical data with the knowledge of traffic flow theory and intelligent transportation systems. His contacts
are as follows
Email: he.zb@hotmail.com,
Webpage: http://zhengbing.weebly.com,
Researchgate page: https://www.researchgate.net/profile/Zhengbing_He, and
ORCID: http://orcid.org/0000-0001-5716- 3853
August 24, 2017
Please cite the paper in the following way
Zhengbing He, Research based on high-fidelity NGSIM vehicle trajectory datasets: A review,
doi:10.13140/RG.2.2.11429.60643, 2017
Contents
1 Introduction 3
2 NGSIM trajectory datasets 3
3 Selection of literature reviewed 4
4 Subjects of the research based on NGSIM datasets 6
4.1 Microscopic traffic flow analysis and modeling ................. 6
4.1.1 Model-based car-following modeling ................... 7
4.1.2 Date-driven car-following modeling ................... 9
4.1.3 Car-following behavior analysis ..................... 10
4.1.4 Lane-changing modeling ......................... 11
4.1.5 Driving strategy development ...................... 11
4.2 Mesoscopic and macroscopic traffic flow modeling ............... 12
4.3 Traffic-related estimation and prediction .................... 13
4.3.1 Macroscopic traffic variables estimation ................. 13
4.3.2 Macroscopic traffic phenomena estimation and analysis ........ 15
4.3.3 Traffic states estimation ......................... 15
4.3.4 Travel time estimation .......................... 16
4.3.5 Intersection traffic-related estimation .................. 16
4.4 Traffic flow model calibration .......................... 17
4.5 Vehicle trajectory data cleaning ......................... 18
4.6 Vehicular Ad Hoc Network-related studies ................... 19
5 Usage 20
6 Limitations 21
7 Outlook 23
8 Summary 23
2
1. Introduction
With the arrival of the 21st century, “big data” has become one of the hottest words that
we almost hear every day, and various data seem to be everywhere. However, even in such an
era of data explosion, the high-fidelity vehicle trajectory data published in Next Generation
Simulation (NGSIM) program ten years ago (U.S. Federal Highway Administration,2006)
still play a unique role in transportation research, due to the facts (1) the NGSIM datasets
are ones of the few datasets providing high-fidelity vehicle trajectories, although their time-
space coverage is limited; (2) by utilizing the NGSIM datasets, a large amount of research
was carried out, making people understand traffic unprecedentedly deeply.
Therefore, it is of interest and significance to know what research was conducted based
on the NGSIM datasets. Better understanding the past will shed light on the future to
making better use of such data and to further collect useful or complementary data. To
the end, the review introduces the research based on the high-fidelity NGSIM trajectory
datasets, by classifying the research and pointing out the main usage, limitations of the
NGSIM datasets, and outlooks of future data collection. Different from most of literature
reviews that organized papers along one subject stream, the review summarizes the papers
that are all linked by the same data. Moreover, to the best of the authors’ knowledge, this is
the first time that the research based on high-fidelity vehicle trajectory data or the NGSIM
datasets is specially reviewed.
The rest of the review is organized as follows: Section 2briefly introduces the NGSIM
datasets including the US-101, I-80, Peachtree, and Lankershim datasets; Section 3presents
the rule of selecting the papers that are reviewed here; Section 4classifies and reviews those
papers; Sections 5-7point out the main usage, limitations of the NGSIM datasets, and
outlooks of future data collection, respectively; Sections 8gives a summary to close the
review.
2. NGSIM trajectory datasets
The NGSIM datasets were originally collected by using cameras, and then extracted from
the resulting videos. The sampling frequency of the NGSIM trajectory is 0.1 sec, and each
sample includes the information such as instantaneous speed, acceleration, longitudinal and
lateral positions, vehicle length, vehicle type. The descriptions of the four datasets are given
as follows.
The US-101 trajectory dataset was collected on a segment in the vicinity of Lankershim
Avenue on southbound US-101 freeway in Los Angeles, California. The segment is approx-
imately 640 m in length, and contains 6 lanes (see Figure 1(a)). The time period of data
collection is 45 min, i.e., from 7:50 a.m. to 8:35 a.m. on June 15, 2005.
3
The I-80 trajectory dataset was collected on a segment of I-80 freeway in Emeryville (San
Francisco), California. The segment is approximately 500 m in length, and contains 6 lanes,
where the median lane is a high occupancy vehicle (HOV) lane (see Figure 1(b)). The data
were collected within two periods, i.e., 15 min ranging from 4:00 p.m. to 4:15 p.m. on April
13, 2005, and 30 min ranging from 5:00 p.m. to 5:30 p.m. on April 13, 2005.
The Peachtree trajectory dataset was collected on a segment of Peachtree Street in At-
lanta, Georgia. The arterial segment is approximately 640 m in length, with five intersections
(four are signalized and one is not) and two or three through lanes in each direction (see
Figure 1(c)). The dataset consists of two 15-min time periods, 12:45 p.m. to 1:00 p.m. and
4:00 p.m. to 4:15 p.m. on November 8th, 2006.
The Lankershim trajectory dataset was collected on a segment of Lankershim Boulevard
in the Universal City neighborhood of Los Angeles, California. The segment is approximately
488 m in length, and contains three or four lanes and four signalized intersections (see Figure
1(d)). The time period of data collection is 30 min ranging from 8:30 a.m. to 9:00 a.m on
June 16, 2005.
3. Selection of literature reviewed
Initially, we searched for a keyword of “NGSIM” in Google Scholar, and obtained over
1700 results2. Then, searched in the homepages of mainstream transportation journals,
and obtained total 227 results; see Table 1. The large number of results implies that the
popularity and importance of the NGSIM datasets, whereas it makes reviewing all of them
difficult. Therefore, to filter the results, we only review the literature published in the
journal series from Transportation Research Part A (TR-A) to Transportation Research
Part F (TR-F). It is known that the Transportation Research series published by Elsevier
are leading journals almost covering all aspects of transportation research. It is believed that
the literature published in the series is able to well represent the state of the art, although
it is admitted that a large number of important publications may be missed3.
After carefully reading those papers, those only mentioning “NGSIM” (e.g. when intro-
ducing other works or pointing out future research direction) instead of actually using the
datasets were removed. Finally, total 71 papers were reviewed here4; see Table 1for the
numbers of the reviewed papers in each journal. It is noticed that most of the papers were
2Google scholar: http://scholar.google.com. It was searched on March 20, 2017
3Considering the fact that the scope of IEEE Transactions on Intelligent Transportation Systems is similar
to that of TR-C (Lijun and Yafeng,2017), the 48 papers published in the transaction are not reviewed here,
whereas the generality of the following subject classifications may not lose much.
4Note that we only briefly introduce the contribution of the literature, since the purpose of this review
is to help readers understand what kind of research was conducted based on the NGSIM datasets. For the
detailed model or method, interesting readers may refer to the original papers.
4
(d) (c)
(a)
(b)
Figure 1: Schematic diagrams of sites collecting the NGSIM trajectories: (a) US-101; (b) I-80; (c) Peachtree;
(d) Lankershim.
5
Table 1: The results of searching for “NGSIM” in transportation journals
Journals Number of Number of
all papers reviewed papers
Transportation Research Part A: Policy and Practice 0 0
Transportation Research Part B: Methodological 52 40
Transportation Research Part C: Emerging Technologies 56 30
Transportation Research Part D: Transport and Environment 4 0
Transportation Research Part E: Logistics and Transportation Review 0 0
Transportation Research Part F: Traffic Psychology and Behaviour 2 1
IEEE Transactions on Intelligent Transportation Systems 48 —
Computer-Aided Civil and Infrastructure Engineering 7 —
IET Intelligent Transport Systems 8 —
Journal of Intelligent Transportation Systems 10 —
Transportmetrica A: Transport Science 11 —
Transportmetrica B: Transport Dynamics 4 —
ASCE Journal of Transportation Engineering 14 —
Transport Policy 2 —
Networks and Spatial Economics 2 —
Journal of Transport Geography 0 —
Transportation 0 —
Accident Analysis and Prevention 7 —
Total 227 71
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017.3
0
2
4
6
8
10
12
14
16
18
Years
Number of publications
Figure 2: The number of the 71 reviewed papers in each year.
published in TR-B and TR-C, implying that high-fidelity trajectory data were more widely
used in (also may be more useful for) the studies of transportation-related methodology and
technology instead of transportation-related policy (TR-A), environment (TR-D), logistics
(TR-E) and psychology (TR-F). Figure 2presents the trend of the publications in recent
years, indicating that although more than ten years have passed the NGSIM datasets are
still popularly used.
4. Subjects of the research based on NGSIM datasets
4.1. Microscopic traffic flow analysis and modeling
Total 25 papers are found under the subject of microscopic traffic flow analysis and
modeling, which includes five sub-subjects as follows: model-based car-following modeling,
data-driven car-following modeling,car-following behavior analysis,lane-changing modeling,
6
and driving strategy development (refer to Table 2). It is not difficult to understand that
model-based car-following modeling and car-following behavior analysis, which take the ad-
vantage of the high resolution of the NGSIM datasets, are two main sub-subjects that include
eight and seven papers, respectively. The specific review on each paper and the respective
data usage are given as follows.
4.1.1. Model-based car-following modeling
Formulating car-following behaviors in mathematics is called model-based car-following
modeling, which is the most prevailing way of modeling driving behaviors. Tordeux et al.
(2010) proposed an adaptive time gap car-following model, which is able to produce the
leader-follower interaction by adjusting the time gap to a targeted safety time, i.e., a function
of speed. The US-101 dataset was used to estimate the distribution of the time gaps under
different vehicle speeds, i.e., the model parameters for each class of vehicles. Koutsopoulos
and Farah (2012) presented a flexible framework to model car-following behavior, which is
based on the recognition of driving regimes, such as car-following, free-flow, emergency stop-
ping. In each regime, driver’s decisions, such as acceleration, deceleration, and do-nothing,
could be made depending on surrounding traffic conditions. Selected leader-follower pairs
from the I-80 dataset were used to calibrate both the proposed model and a modified Gen-
eral Motors car-following model (Ahmed,1999), and those two models were compared using
Akaike’s information criterion (Akaike,1974). To reproduce traffic oscillations, Chen et al.
(2012) proposed a behavioral car-following model, which is based on the empirical finding
from the US-101 trajectories that driver behavior before an oscillation is strongly correlated
to that during the oscillation. The US-101 trajectories were employed to graphically and
statistically analyze car-following behavior, and to calibrate the model parameters. Laval
et al. (2014) proposed a desired acceleration model and incorporated it into the framework of
Newell’s simplified car-following model (Newell,2002). To validate the model, stop-and-go
waves were simulated. The simulated pattern of oscillation growth and hysteresis were com-
pared with empirical ones, and a lead vehicle problem (LVP) for 6th follower5was tested.
After observing the heterogeneity of driving behavior in different spacing-speed states from
the US-101 trajectories, Tian et al. (2015) assumed oscillating spacing and incorporated it
into a cellular automaton to reproduce the empirical observations of Kerner’s three-phase
theory (Kerner,2009).
5Lead vehicle problem for the nth follower: given leader’s complete trajectory and nfollowers’ initial
positions and speeds, simulate those followers, and compare the profiles (such as speed and spacing) of the
simulated nth follower with those of real nth follower.
7
Table 2: General information of the papers regarding microscopic traffic flow analysis and modeling
Sub-subjects Papers Dataset Journals Main usage of data
Model-based Tordeux et al. (2010) US-101 TR-B Estimating empirical distribution of time gap and model parameters
car-following Koutsopoulos and Farah (2012) I-80 TR-B Calibrating model and comparing with other models
modeling Chen et al. (2012) US-101 TR-B Graphically and statistically analyzing car-following behavior and calibrating model
Laval et al. (2014) US-101 TR-B Extracting macroscopic traffic patterns as references and testing an LVP for 6th follower
Tian et al. (2015) US-101 TR-B Analyzing trajectories for typical driving characteristics
Wang et al. (2011) I-80 TR-C Testing an LVP for 1st follower
Przybyla et al. (2015) US-101, I-80 TR-C Training model and comparing with simulated trajectories
Hamdar et al. (2015) I-80 TR-B Calibrating model using a nonlinear optimization procedure based on a generic algorithm
Validating model through a test similar to 2-fold cross-validation
Data-driven Zheng et al. (2013) US-101 TR-C Training model (70% data) and validating model (30% data, and an LVP for 8th follower)
car-following He et al. (2015) US-101, I-80 TR-B US-101: training model, testing LVPs for 23rd and 30th followers, and
modeling extracting macroscopic traffic patterns as references; I-80: showing transferability
Hao et al. (2016) US-101 TR-C Three trajectories for model calibration, and four tra jectories to test an LVP for 1st follower
Car-following Chiabaut et al. (2010) I-80 TR-B Graphically and statistically analyzing leader-follower trajectories
behavior Chen et al. (2012) US-101 TR-B Graphically and statistically analyzing leader-follower trajectories
analysis Chen et al. (2014) US-101 TR-B Graphically and statistically analyzing trajectories
Wei and Liu (2013) US-101 TR-B 5-fold cross-validation and an LVP for 1st follower
Li et al. (2013) US-101 TR-F Estimating headway distributions and model parameters
Taylor et al. (2015) I-80 TR-B Estimating the distributions of model parameters
Hamdar et al. (2016) I-80 TR-C Building simulation environments with empirical traffic
Lane-changing Laval and Leclercq (2008) I-80 TR-B Illustrating traffic phenomena
modeling Zheng et al. (2013) US-101, I-80 TR-B Illustrating lane-changing process, calibrating and validating models
Talebpour et al. (2015) US-101 TR-C Calibrating and validating lane-changing model
Balal et al. (2016) US-101, I-80 TR-C I-80: calibrating and validating model; US-101: testing for transferability
Driving Yang and Jin (2014) US-101 TR-C Building simulation environments with an empirical leader and an initial condition
strategy Gong et al. (2016) I-80 TR-C Building a simulation environment with an empirical leader
development Tak et al. (2016) US-101 TR-C Providing macroscopic and microscopic data, calculating collision risk,
and analyzing system performance
8
Three safety-related car-following studies are founded and introduced as follows. To un-
derstand vehicle-to-vehicle dynamic interactions and prevent rear-end collisions, Wang et al.
(2011) proposed a driver’s safety approaching behavioral model by considering the variability
of a follower’s speed and spacing to its leader. To validate the model, an empirical leader
was selected from the I-80 dataset, and the LVP was tested for 1st follower. To estimate
risk effects of driving distraction, Przybyla et al. (2015) proposed a dynamic errorable car-
following model by using a dynamic time warping algorithm. Leader-follower car-following
pairs were extracted from both the US-101 and I-80 trajectory datasets and used to train
the dynamic time warping algorithm. Subsequently, the simulated trajectories were com-
pared with real ones to analyze the risk effects. Hamdar et al. (2015) proposed a stochastic
car-following model with some important behavioral and psychological considerations, such
as subjective utilities and dis-utilities for acceleration and deceleration, risk taking. To cali-
brate the stochastic model, a nonlinear optimization procedure based on a generic algorithm
(Hamdar et al.,2009) was used. A test analog to 2-fold cross-validation6was conducted by
splitting the I-80 dataset into two folders, i.e., one is between 4:00 p.m. to 4:15 p.m., and
the other is between 5:00 p.m. to 5:15 p.m.
4.1.2. Date-driven car-following modeling
Compared with the model-based modeling, date-driven car-following modeling is mainly
based on artificial intelligence without mathematics directly describing car-following behav-
iors. Zheng et al. (2013) developed a neural-network-based car-following model by building
and incorporating a neural network for instantaneous reaction delay. 70% of the US-101
data were used to train the back-propagation algorithm, and the remaining 30% were used
to test an LVP for 8th follower. He et al. (2015) proposed a K-Nearest-Neighbor based
car-following model, whose four inputs are leader’s moving distances and follower’s space
headways in the latest two time steps, and whose output is follower’s moving distance. The
US-101 dataset was used to train the model, and the I-80 dataset was compared to show
the transferability of the proposed model. To validate the model, LVPs for 23rd and 30th
followers were tested, and important traffic characteristics such as wave speed, fundamen-
tal diagrams were extracted from the US-101 dataset as references to compare with the
simulated ones. Hao et al. (2016) proposed a fuzzy logic-based car-following model with a
five-layer structure, i.e., Perception-Anticipation-Inference-Strategy-Action. Seven trajecto-
ries were selected from the US-101 trajectories on the leftmost lane, three of which were used
6k-fold cross-validation: the original sample is randomly partitioned into kequal sized folders. Of the k
folders, a single folder samples is retained as the validation data for testing the model, and the remaining
m−1 folder samples are used as training data. The cross-validation process is repeated ktimes (the folds),
with each of the kfolders used exactly once as the validation data. The kresults from the folds can then be
averaged to produce a single estimation.
9
for model calibration, and four of which were used for validation by testing an LVP for 1st
follower.
4.1.3. Car-following behavior analysis
Comparing with car-following modeling, the papers in this sub-subject more focus on
analyzing car-following behaviors and finding microscopic driving characteristics. Chiabaut
et al. (2010) carefully analyzed the parameters of Newell’s car-following model (Newell,2002)
in congested traffic conditions, and then established relations between stochastic Newell’s
model with heterogeneous drivers and its associated macroscopic pattern. Chen et al. (2012)
studied traffic hysteresis from a behavioral perspective, and found that the occurrence and
type of traffic hysteresis is closely correlated with driver behavior when experiencing traffic
oscillations. Moreover, Chen et al. (2014) further investigated on the formation of traffic
phenomena from the car-following perspective, including how periodic oscillation forms, how
driver characteristics contribute to capacity drop, etc.
All the literature carried on their studies mainly by analyzing the leader-follower tra-
jectories graphically and statistically, in particular on the basis of Newell’s car-following
theory (Newell,2002), such as estimating the distribution of wave speed, plotting time-space
trajectory and flow-density diagrams.
Wei and Liu (2013) employed a self-learning support vector regression approach to in-
vestigate the asymmetric characteristic in car-following as well as its impacts on traffic flow
evolution. 5-fold cross-validation was conducted to train and validate the approach, and an
LVP for 1st follower was tested for model validation. Li et al. (2013) proposed an asymmetric
stochastic extension of the Tau Theory (Lee,1976) to explain the phenomenon that vehicle’s
headway follows a certain log-normal type distribution within different speed ranges, imply-
ing that the physiological Tau characteristics implicitly affect human driving behavior and
the resulting traffic dynamics. The US-101 dataset was used to estimate the distributions
of headways as well as the parameters of the Tau Theory. To examine intradriver hetero-
geneity, Taylor et al. (2015) proposed a dynamic time warping algorithm to analyze vehicle
trajectories and car-following behavior. The proposed dynamic time warping algorithm was
employed to extract the distributions of the parameters in a stochastic extension of Newell’s
car-following model (Newell,2002) from the I-80 dataset. To quantify driver behavior under
different roadway geometries and weather conditions, Hamdar et al. (2016) extended the
acceleration modeling framework based on the Prospect Theory (Kaneman and Tversky,
1979), and conducted driving experiments using a driving simulator. Typical I-80 trajecto-
ries were selected to build the simulated traffic environment (at least five vehicles were in
front of and behind the lead vehicle, respectively), in order to make it as generic as possible.
10
4.1.4. Lane-changing modeling
Lane-changing is more complicated driving behavior including simultaneous longitudinal
and lateral movements. Four papers are founded to model the lane-changing behavior in
a microscopic perspective. Laval and Leclercq (2008) introduced a framework to capture
the relaxation phenomena commonly observed near congested on-ramps, i.e., vehicles take
short spacings when entering a freeway, but “relax” to more comfortable spacings shortly
thereafter. The I-80 data were plotted in time-space trajectory and flow-density diagrams to
illustrate the relaxation phenomena. Zheng et al. (2013) investigated the anticipation and re-
laxation processes in lane-changing, and offered an extension to Newell’s car-following model
(Newell,2002) to describe a regressive effect. The trajectories of lane-changers were selected
and graphically analyzed, and the parameters (representing anticipation and relaxation) in
the extension of Newell’s car-following model were calibrated and statistically investigated.
To validate the model, the selected lane-changing data were split into two subsets, and
their modeling results were compared. Talebpour et al. (2015) proposed a lane-changing
model in a connected-vehicle environment, based on a game theory (two-person non-zero-
sum non-cooperative game) approach that endogenously accounts for the information flow.
Mandatory and discretionary lane-changing behaviors were identified in the US-101 dataset.
15 min (7:50 a.m. to 8:05 a.m.) data were used to calibrate the model, and another 15
min (8:05 a.m. to 8:20 a.m.) data were used to validate it by calculating estimation errors.
Based on a fuzzy inference system, Balal et al. (2016) presented a binary decision model to
determine if it is time to execute a lane change. Randomly-selected 70% trajectories of the
I-80 dataset were used to calibrate the fuzzy inference system, and the remaining 30% were
used to test the results. In addition, the US-101 dataset was served to test the calibrated
model in order to show its transferability.
4.1.5. Driving strategy development
To reduce driving emissions or to increase driving safety, driving strategies or driver
assistance systems utilizing new technologies are proposed in the following three papers.
Yang and Jin (2014) developed a distributed cooperative green driving strategy based on
inter-vehicle communications, in order to smooth traffic flow and lower pollutant emissions
and fuel consumption in stop-and-go traffic. Ten trajectories with a high standard deviation
in speed were selected and combined as a leading trajectory, and the speed profiles of the
follower before and after applying the proposed strategy were compared to demonstrate
the effect of the proposed strategy. In addition, the initial locations of the vehicles in
the US-101 dataset were taken as an initial condition to build a simulation scenario to
test the effect of the strategy with different penetration rates and communication delays.
Gong et al. (2016) proposed a car-following control scheme for a platoon of connected and
autonomous vehicles, which was modeled as an interconnected dynamic system subject to
11
acceleration, speed, and safety distance constraints. An oscillating vehicle trajectory was
selected from the I-80 dataset and taken as an empirical leader of a simulation scenario.
Tak et al. (2016) developed a hybrid collision warning system that was able to utilize the
integrated information of macroscopic loop detector data and microscopic smartphone data.
The US-101 trajectories were taken as the microscopic data, and virtual loop detectors were
set to calculate macroscopic data. The collision risk resulting from three collision warning
systems was calculated and compared, and two detailed vehicle trajectories were selected to
analyze the performance of the systems.
4.2. Mesoscopic and macroscopic traffic flow modeling
Only five papers are found under the subject of mesoscopic and macroscopic traffic flow
modeling (refer to Table 3). Chiu et al. (2010) proposed a vehicle-based mesoscopic traf-
fic simulation model that explicitly considers the anisotropic property of traffic flow (i.e.,
vehicles mostly react to other vehicles that are in front of them (Daganzo,1995)) into the
vehicle state update at each simulation step. The I-80 dataset was employed to calibrate
the proposed model by minimizing the introduced variable of a speed influencing region.
Piccoli et al. (2015) presented several data fusion schemes to incorporate vehicle trajectory
data into a second-order phase transition model, and evaluated the estimation accuracy of
first-order variables (such as speed and density) and second-order variables (such as accelera-
tion and emission), by using mobile sensor data with various penetration rates and sampling
frequency. A standard k-fold cross-validation was conducted, and the value of kwas set
based on the penetration rates. Qian et al. (2017) developed a macroscopic heterogeneous
traffic flow model by considering interplay of multiple vehicle classes, and introduced an
intuitive computational procedure to capture mixed vehicular flow propagation and shock
wave formation. To validate the model, an initial value problem7(IVP) based on the I-80
dataset was solved.
Two papers are related to the macroscopic traffic models taking lane-changing into ac-
count. Jin (2010) proposed a macroscopic kinematic wave model to capture bottleneck effects
and to aggregate traffic dynamics of lane-changing traffic. Jin (2013) developed a multi-
commodity (i.e., weaving and non-weaving vehicles) behavior Lighthill-Whitham-Richards
model of lane-changing traffic flow. A lane-changing fundamental diagram was introduced,
which is determined by both car-following and lane-changing characteristics as well as road
geometry and traffic composition. Both the studies calibrated the newly introduced concepts,
i.e., the lane-changing intensity and the lane-changing fundamental diagram, by measuring
7Initial value problem: given empirical boundary and initial traffic conditions extracted from empirical
data, estimate intermediate traffic states using the proposed model and then compare them with the real
traffic states.
12
empirical traffic variables from the I-80 dataset, such as on-ramp flow rate, lane-changing
times.
Table 3: General information of the papers regarding mesoscopic and macroscopic traffic flow modeling
Papers Dataset Journals Main usage of data
Chiu et al. (2010) I-80 TR-B Calibrating model by formulating an optimization problem
Piccoli et al. (2015) I-80 TR-C k-fold cross-validation
Qian et al. (2017) I-80 TR-B Validating model: inputting as boundary and initial conditions of IVP
Jin (2010) I-80 TR-B Calibrating model, i.e., the lane-changing intensity
Jin (2013) I-80 TR-B Calibrating model, i.e., the lane-changing fundamental diagram
4.3. Traffic-related estimation and prediction
The subject of traffic-related estimation and prediction is another main subject in utilizing
the NGSIM datasets. Total 26 papers are found under the subject, and five sub-subjects are
further classified as presented in Table 4and introduced as follows.
4.3.1. Macroscopic traffic variables estimation
In this sub-subject, macroscopic traffic variables, such as traffic flow, speed, are estimated
based on the NGSIM datasets, and the fundamental diagrams are constructed. Coifman
(2015) presented a method to estimate the fundamental diagrams with no need of seek out
stationary conditions. Vehicle length was found to be the key. As a complement of loop
detector data, the I-80 dataset with higher resolution was analyzed by using the proposed
method to increase the credibility of the method. Siqueira et al. (2016) proposed an alter-
native stochastic model for the fundamental diagrams, by introducing a stochastic transport
model with discrete speed spectrum. The I-80 dataset was used to calculate model param-
eters and to estimate the empirical fundamental diagrams, which were then compared with
the estimation results made by the proposed model. Based on Newell’s car-following model
(Newell,2002), Jabari et al. (2014) proposed a stochastic version of the macroscopic traffic
flow speed-density relation, which allows to investigate the impact of driver heterogeneity on
macroscopic traffic flow relations. The first 15-min (4:00 p.m. to 4:15 p.m.) I-80 data were
used to estimate the distributions of model parameters, and the other 30-min (5:00 p.m. to
5:30 p.m.) data were used to plot an empirical speed distribution, which was then compared
with the simulation results. Wu and Coifman (2014) proposed a length-based vehicle clas-
sification method from dual-loop detectors by considering vehicle acceleration in congested
traffic. Through setting virtual loop detectors, the I-80 dataset was used to evaluate the
performance of the proposed vehicle classification method. The vehicle length information
contained in the high-fidelity NGSIM datasets makes evaluating the vehicle length-based
study possible.
13
Table 4: General information of the papers regarding macroscopic traffic variables estimation
Sub-subjects Papers Dataset Journals Main usage of data
Macroscopic Coifman (2015) I-80 TR-B Being analyzed by using the proposed method as a complement of loop detector data
traffic variables Siqueira et al. (2016) I-80 TR-B Calibrating model parameters and estimating the referred fundamental diagram
estimation Jabari et al. (2014) I-80 TR-B Calibrating the distributions of model parameters using the (first 15 min) I-80 data,
and validating model by comparing with speed distributions (other 30 min)
Wu and Coifman (2014) I-80 TR-C Evaluating performance of the proposed method after setting virtual loop detectors
Macroscopic Laval (2011) US-101 TR-B
traffic Ahn et al. (2013) US-101, I-80 TR-C
phenomena Zheng et al. (2011a) US-101 TR-B Applying the respective methods to extract and measure macroscopic traffic
estimation and Zheng et al. (2011b) US-101, I-80 TR-B phenomena or the phenomena-related variables
analysis Blandin et al. (2013) I-80 TR-B
Oh and Yeo (2015) US-101 TR-B
Li et al. (2014) US-101 TR-B
Traffic states Herrera and Bayen (2010) US-101 TR-B Calculating ground-truth traffic states and simulating various probe vehicle data
estimation Deng et al. (2013) I-80 TR-B Calculating ground-truth traffic states
Bucknell and Herrera (2014) US-101 TR-C Calculating ground-truth traffic states and simulating various probe vehicle data
Argote-Caba˜nero et al. (2015) Peachtree TR-C Calculating ground-truth measures of effectiveness and simulating connected vehicle data
Travel time Ramezani and Geroliminis (2012) Peachtree TR-B Calculating ground-truth travel time for evaluating estimation results
estimation Feng et al. (2014) Peachtree TR-C
Intersection Qi et al. (2013) Lankershim TR-C Calibrating the triangular fundamental diagram and the proposed model
traffic-related Srivastava et al. (2015) Lankershim TR-B Calibrating the triangular fundamental diagram
estimation Sun and Ban (2013) Peachtree TR-C
Hao et al. (2013) Peachtree TR-C
Hao et al. (2014) Peachtree TR-B Validating model, since the data reflect complete ground-truth traffic
Sun et al. (2013) Peachtree TR-B
Yang et al. (2016) Lankershim TR-C
Lee et al. (2015) Lankershim TR-C Calibrating model: first 10-cycle data; validating model: remaining 10-cycle data
Lee and Wong (2017) Lankershim TR-B Validating model
14
4.3.2. Macroscopic traffic phenomena estimation and analysis
This sub-subject studies on macroscopic traffic phenomena, such as hysteresis, oscilla-
tions, capacity drop, based on the NGSIM datasets. Laval (2011) found a new shape for the
well-known hysteresis phenomenon in traffic flow after aggregating time-space trajectories
along the wave direction by using Edie’s definitions (Edie,1963). Ahn et al. (2013) later
investigated the hysteresis as vehicles experienced stop-and-go waves, and it was found that
the hysteresis takes place less frequently and in smaller amplitude than previously thought.
Zheng et al. (2011a) applied wavelet transform to analyze important features related to bot-
tleneck activations and traffic oscillations in congested traffic in a systematic manner. Zheng
et al. (2011b) subsequently demonstrated a way of using wavelet transform to identify the
formation and propagation of the stop-and-go waves. Blandin et al. (2013) presented a phase
transition model of non-stationary traffic to model complex macroscopic traffic phenomena
such as hysteresis and phantom jams. Oh and Yeo (2015) analyzed the capacity drop from a
microscopic perspective, and found several factors that may trigger the drop, such as driver’s
tendency to take a large headway after passing stop-and-go waves. Using their previously
proposed describing-function based approach (Li et al.,2012), Li et al. (2014) estimated fuel
consumption and emission from traffic oscillations and explored vehicle control strategies to
smooth traffic.
All the above studies regarding macroscopic traffic phenomena estimation and analysis
mainly applied the respective methods to extract and measure macroscopic traffic phenomena
or the phenomena-related variables, such as traffic flow, capacity, and headway, from the
NGSIM trajectories.
4.3.3. Traffic states estimation
This sub-subject focuses on estimating traffic states, in particularly estimating inter-
mediate traffic states given downstream and upstream boundaries. Given downstream and
upstream traffic states to estimate intermediate states, Herrera and Bayen (2010) presented
a method to incorporate mobile probe sensor data into a freeway traffic flow model. This
method can work even when data are not available for the on- and off-ramps. Deng et al.
(2013) extended Newell’s three-detector problem (Newell,1993), and presented a stochas-
tic traffic state estimation method using multiple data sources, such as loop detector and
floating car data. Bucknell and Herrera (2014) studied on the accuracy of traffic states
estimation (especially in reconstructing traffic speed) using various penetration rates and
sampling frequency of mobile sensors. Argote-Caba˜nero et al. (2015) studied on the estima-
tion of measures of effectiveness (such as average speed, number of stops, delay) for traffic
operations under a connected-vehicle environment, and determined the minimum connected-
vehicle penetration rate to estimate the measures of effectiveness. Since the Peachtree dataset
is lack of saturated conditions, only undersaturated conditions were studied using the em-
15
pirical data in Argote-Caba˜nero et al. (2015).
Taking the advantage that the NGSIM datasets provide a complete picture of the mon-
itored traffic, the papers under this sub-subject employed the NGSIM datasets to calculate
ground-truth traffic states or variables for model or method evaluation, or to simulate probe
vehicle data or connected vehicle data with different penetration rates and/or sampling fre-
quency.
4.3.4. Travel time estimation
Two papers studying on travel time estimation are found. Given probe vehicles’ travel
times, Ramezani and Geroliminis (2012) estimated arterial travel time by applying a Markov
chain procedure to integrate travel time correlation of routes successive links. To estimate
arterial travel time using probe vehicle data, Feng et al. (2014) employed mixtures of normal
distributions to approximate link travel time distributions, and then estimated real-time
travel time using probe vehicle travel time based on the Bayes Theory.
Similar to the studies of estimating traffic states, the data usage here is also to provide
ground-truth travel time for evaluating estimation results.
4.3.5. Intersection traffic-related estimation
All the papers focusing on estimating traffic states or variables at intersection, such as
discharge flow, queue length at intersection, are placed in this sub-subject. Qi et al. (2013)
established the relationship between discharge flow and (efficiency- and objective-driven)
lane-changing behavior at signalized intersections, and proposed an enhanced Cell Transmis-
sion Model by incorporating the lane-changing behavior. To model the realistic discharge
flow rate and headway features at signalized intersections, Srivastava et al. (2015) presented
a modified Cell Transmission Model by substituting the traditional demand function with a
linearly decreasing function, and solved it under various Riemann problem scenarios.
Since both of the above studies were based on the Cell Transmission Model and focused on
the traffic at signalized intersections, the Lankershim dataset was thus employed to calibrate
the triangular fundamental diagram simply by a regression method or observation. In Qi
et al. (2013), a driving behavioral parameter in the proposed model was also calibrated.
Sun and Ban (2013) presented optimization- and delay-based models based on a vari-
ational formulation of traffic flow (Daganzo,2005) to reconstruct vehicle trajectories from
mobile traffic sensors at arterial intersections. Hao et al. (2013) proposed a three-layer
Bayesian Network model to describe the stochastic intersection flow, by capturing the rela-
tionship between the arrival and departure processes and vehicle indices, which is a newly
introduced concept in a cycle at a signalized intersection. Hao et al. (2014) presented a
Bayesian Network based model to estimate the cycle-by-cycle queue length distribution of
a signalized intersection, by using sample travel times collected from mobile sensors. To
16
balance data needs for transportation modeling and privacy protection, Sun et al. (2013)
developed a virtual trip lines zone-based system that is able to ensure an acceptable level of
privacy and result in satisfactory results of transportation applications. Yang et al. (2016)
improved a signal control algorithm developed for connected vehicles (Ilgin Guler et al.,
2014) in several aspects, such as integrating three different stages of technology develop-
ment, developing a heuristic method to switch the signal controls, incorporating trajectory
design for automated vehicles.
All the above studies are based on partial traffic data (such as collected by mobile sensors
or in virtual trip lines zones), and thus the advantage of the NGSIM trajectory data, i.e.,
providing complete traffic information, is taken to validate the proposed estimation methods,
which is similar to the data usage in the estimations of traffic states and travel time. It is
worth mentioning that to overcome the insufficiency issue of the NGSIM datasets, Hao et al.
(2014) ran 20 replicas of the estimation method for each cycle under each penetration rate.
To estimate lane-based queue lengths at intersection in real time, Lee et al. (2015) devel-
oped discriminant models to identify critical issues, such as where there is a residual queue
and estimating downstream arrivals for each lane. Lee and Wong (2017) proposed a group-
based approach to estimate lane-based incremental queue accumulations, and presented a
control delay model to predict temporal and spatial factors in future incremental queue ac-
cumulations as well as to produce the most appropriate time windows for a rolling horizon
procedure.
Both the above papers conducted their lane-based studies by utilizing the high resolution
of the NGSIM datasets collected at intersection, i.e., extracting lane-based queue lengths.
In addition, Lee et al. (2015) converted the Lankershim trajectories to loop detector data by
setting virtual downstream and upstream loop detectors. Then, the data collected during
the first 10 cycles were used to calculate model parameters, and the remaining 10-cycle data
were used to validate the model by comparing empirical and simulated queue lengths.
4.4. Traffic flow model calibration
Making use of the NGSIM datasets, total eight papers study on the traffic flow model
calibration as shown in Table 5.
Table 5: General information of the papers regarding traffic flow model calibration
Papers Dataset Journals Main usage of data
Li et al. (2012) US-101 TR-B
Rhoades et al. (2016) US-101 TR-B
Kim et al. (2013) US-101 TR-C
Vieira da Rocha et al. (2015) I-80 TR-D Being an empirical sample to
Li et al. (2016) US-101 TR-C demonstrate the proposed method
Durrani et al. (2016) US-101 TR-C
Zhong et al. (2016) I-80 TR-C
Sopasakis and Katsoulakis (2016) US-101 TR-B
17
By using vehicle trajectory data with extracted frequency-domain characteristics, Li et al.
(2012) proposed a systematic framework to validate the describing-function approach (Li
et al.,2012) that is an analytical approach able to predict traffic oscillation propagation for
a general class of car-following models. To calibrate nonlinear car-following laws based on
leader-follower trajectories, Rhoades et al. (2016) proposed a calibration method that takes
into account not only driver’s car-following behavior but also time- and frequency-domain
properties of vehicle trajectories. Kim et al. (2013) proposed a robust algorithm called
the expectation-maximization to calibrate a General Motors car-following model (Chandler
et al.,1958) with random coefficients reflecting drivers’ heterogeneity. Vieira da Rocha et al.
(2015) studied on the effectiveness of goodness-of-fit indicator-based calibration method of
car-following models in estimating environment-related factors, such as fuel consumption,
nitrogen oxide and particulate matter emissions. To better calibrate car-following models,
Li et al. (2016) proposed a global optimization algorithm that integrates global direct search
and local gradient search to find the optimal solution in an efficient manner. Durrani et al.
(2016) calibrated the driving behavior parameters for cars and heavy vehicles in the Wiede-
mann 99 vehicle-following model (Aghabayk et al.,2013), and demonstrated the significant
effect of the leader class on the flower’s behavior in car-following model calibration. Zhong
et al. (2016) proposed a cross-entropy calibration method to identify parameters of determin-
istic car-following models, by formulating it as a stochastic optimization problem. Moreover,
a probabilistic sensitivity analysis algorithm was introduced to identify the most important
parameters to simply the calibration process. Sopasakis and Katsoulakis (2016) proposed a
dynamic model parameterization approach to appraise traffic flow models and to optimize
their performance against time-series traffic data and prevailing conditions by analyzing per-
turbations of vehicle trajectories. A mathematical method that quantifies traffic information
loss was additionally presented.
All the studies regarding traffic flow model calibration mainly took the NGSIM dataset
as an sample to apply and demonstrate their proposed methods.
4.5. Vehicle trajectory data cleaning
As early pointed out in Punzo et al. (2011), there are errors in the NGSIM datasets, in
particular in the instantaneous data of speed and acceleration. Therefore, five papers aiming
at (trajectory) data cleaning are found as presented in Table 6and as introduced as follows.
To inspect trajectory data accuracy and reduce noise, Punzo et al. (2011) designed quan-
titative methods including jerk analysis, consistency analysis, and spectral analysis, and
applied them on the complete four NGSIM datasets. Cleaned NGSIM datasets were subse-
quently published and expected to be a benchmark for trajectory data quality. Zheng and
Washington (2012) studied on the issue of selecting an optimal wavelet to detect irregu-
lar structures and transient phenomena in traffic data, and recommended the Mexican hat
18
Table 6: General information of the papers regarding vehicle trajectory data cleaning
Papers Dataset Journals Main usage of data
Punzo et al. (2011) US-101, I-80 TR-C
Peachtree, Lankershim
Zheng and Washington (2012) US-101 TR-C Being an empirical sample to
Montanino and Punzo (2015) I-80 TR-B demonstrate the proposed method
Fard et al. (2017) I-80 TR-C
Zheng and Su (2016) US-101 TR-B
wavelet for cleaning traffic and vehicular data. Montanino and Punzo (2015) proposed a
traffic-informed method to denoise and reconstruct vehicle trajectories. A simulation-based
framework was proposed to verify that the reconstructed trajectories were closer to the real
ones (that was actually unknown) than the collected ones. For the similar purpose, Fard
et al. (2017) proposed a two-step technique based on wavelet analysis, i.e., first identifying
and modifying outliers, and then eliminating them by applying a wavelet-based filter. To fill
the gap that the types of noise in traffic data were ignored, Zheng and Su (2016) proposed a
compressed sensing theory based algorithm to recovery traffic data with Gaussian measure-
ment noise, partial data missing, and corrupted noise. Moreover, Markov random field and
total variation regularization were used to improve the accuracy of traffic state estimation.
Since there are various errors in the NGSIM trajectory data, the NGSIM trajectories were
naturally taken as study objects in the studies regarding vehicle trajectory data cleaning.
4.6. Vehicular Ad Hoc Network-related studies
Although Vehicular Ad Hoc Network (VANET) has not widely deployed in practice, two
papers are found to be related to VANET (see Table 7).
Table 7: General information of the papers regarding VANET-related studies
Papers Dataset Journals Main usage of data
Baiocchi (2016) US-101, I-80, Peachtree TR-B Being as VANET testbeds
Du et al. (2016) US-101 TR-C
Aiming at the applications based on VANET, Baiocchi (2016) presented an analytical
model of message coverage distance and delivery delay with timer-based dissemination pro-
tocols, which is able to evaluate the trade-off between delay due to timers and covered
distance. To capture information spreading dynamics via VANET, Du et al. (2016) devel-
oped an information-traffic coupled cell transmission model, which discretizes a road segment
into a number of cells, and mathematically captures the inner-cell and inter-cell movements
of information front and tail.
In the studies, the NGSIM datasets were treated as VANET testbeds to evaluate model
performance, by assuming the VANET applications have been practically deployed.
19
5. Usage
In general, the NGSIM datasets have two main advantages for transportation research.
The first one is its high resolution, which allows researchers to investigate very detailed driv-
ing behaviors and to calibrate or estimate microscopic behavioral parameters and variables.
The second one is its completeness in reflecting traffic conditions, which provides researchers
with 100 percent and ground-truth traffic during the collection period and at the collection
locations. Taking the advantages, those studies mainly utilize the NGSIM trajectory data
in the following ways (also refer to Table 2-7):
•Calibrating or training traffic flow models. To calibrate observable traffic parameters,
such as the fundamental diagrams and a distribution of headways, we can directly estimate
them from the time-space trajectories. For unobservable parameters, the calibration is usu-
ally to solve an optimization problem, in which the decision variables are the parameters of
a model (such as a car-following model) and the objective function characterizes the differ-
ence between empirical vehicle movements and their simulated correspondences. In addition,
for the data-driven models without parameters, the NGSIM datasets are used to train the
model. Note that to eliminate the correlation with model validation, it is usual that only a
part of the dataset is employed to calibrate models, and the other part is used to validate
models.
•Validating models. Three prevailing methods of validating models are as follows. (1)
An LVP test for nth follower, in particular for validating car-following models. (2) (k-fold)
cross-validation. The method is usually used for the model that needs not only validation
but also calibration. (3) Direct comparison with ground truth (such as by calculating er-
ror magnitudes), including comparison with ground-truth traffic states, ground-truth traffic
phenomena, etc. The validation method is widely applied in the studies related to mobile
sensors, and the studies of estimating traffic state and travel time. For mobile sensor-related
studies, the NGSIM datasets not only provide ground-truth traffic, but its completeness also
allows researchers to carry on experiments with different penetration rates of mobile sensors.
Likewise, for traffic state and travel time estimations, the completeness allows researchers to
flexibly set or select any part of traffic states or travel time to be model inputs or outputs.
Moreover, comparing with ground-truth traffic phenomena is an effective way to validate
microscopic (car-following and lane-changing) models from a macroscopic perspective.
•Demonstrating driving behaviors or traffic phenomena. Through graphically and statis-
tically demonstrating and analyzing time-space vehicle trajectories, typical driving behaviors
or traffic phenomena could be obtained and correspondingly modeled. The usage is usually
adopted in car-following and lane-changing modeling and behavior analysis, as well as traffic
phenomena illustration before modeling.
•Being analysis samples. The studies in the subjects, such as macroscopic traffic phe-
20
nomena estimation and analysis,traffic flow model calibration, and vehicle trajectory data
cleaning, need samples to demonstrate and apply their proposed methods. The high-fidelity
NGSIM trajectory data could satisfy the demands and play the role.
•Building simulation environments or testbeds. For the studies related to new technology
that has not been practically applied, such as connected vehicles, VANET, safety and green
driving strategies, the tests are usually conducted in simulation scenarios. Therefore, typical
and representative trajectories or platoons can be selected from the real traffic data, and
used to build simulation environments as realistic as possible.
6. Limitations
For current NGSIM trajectory datasets, the following limitations might exist.
•Limitation-1: The time-space scope of the NGSIM trajectories is limited. For the
freeway (US-101 and I-80) datasets, the time coverages are 45 min, and the space coverages
are approximately 640 m and 500 m, respectively. For the arterial (Peachtree and Lanker-
shim) datasets, the time coverages are 30 min, and the space coverages are approximately
640 m (five intersections) and 488 m (four intersections), respectively.
Such time-space coverages are too short, due to the facts that an oscillation could propa-
gate for 5-10 kilometers (Treiber and Kesting,2012), and that a rush hour of traffic usually
lasts for more than one hours. The shortage may be the most serious issue limiting the
applications of the NGSIM datasets, as reported in Herrera and Bayen (2010); Treiber and
Kesting (2012); Bifulco et al. (2013); Blandin et al. (2013); Wu and Liu (2014); Li et al.
(2014); He et al. (2015); Coifman (2015); Oh and Yeo (2015); Jiang et al. (2015). In par-
ticular, more high-fidelity data, which contain not only car-following behavior but also a
large amount of lane-changing behavior, are very meaningful for data-driven modeling of
car-following behavior, or lane-changing behavior that has not been seen yet.
•Limitation-2: The traffic conditions contained in the NGSIM datasets are limited.
The traffic conditions contained in the freeway datasets are congested traffic conditions with
oscillating features, except some pieces of high flow traffic exhibited by the exiting traffic
in the US-101 dataset and the HOV-lane traffic in the I-80 dataset. The limitation of lack
of high-flow traffic are reported by the literature, such as Chiabaut et al. (2010); Li et al.
(2013); Jin et al. (2015); Balal et al. (2016).
In contrast, the traffic contained in the arterial datasets is unsaturated without residual
queues (Sun and Ban,2013;Argote-Caba˜nero et al.,2015). Therefore, some literature has
to evaluate the model performance under saturated conditions in a simulation environment
instead of using the arterial datasets. Lack of saturated conditions is also a main reason that
21
the studies based on the arterial datasets are less than those based on the freeway datasets8.
•Limitation-3: The variety of data-collection roads is limited. The NGSIM datasets
were collected only from freeways and arterials in U.S., making impossible to investigate
the traffic on other level or type of roads or on the roads in other countries . For example,
urban freeways in some countries, such as China, play a leading role in their road networks.
Compared with freeways, urban freeways have different geometry features, such as denser
and shorter ramps, and lower speed limits (60 km/h-80 km/h). However, due to lack of
high-fidelity datasets such as the NGSIM datasets, some unique traffic phenomena (e.g., the
slowly moving but barely jammed congestion, high-speed car-following behavior) and their
formation mechanisms are still unclear (Guan and He,2008;He et al.,2017).
Moreover, it is a question that to what extent the traffic phenomena or characteristics
founded in the NGSIM datasets could be treated as a universal law; in other words, to what
extent the NGSIM datasets could represent general traffic. Probably, only when more high-
fidelity data are collected on various roads and in different locations, the question could be
properly answered.
•Limitation-4: The variety of traffic components is limited. In general, vehicles can
be classified in three types, i.e., motorcycles, passenger vehicles, and trucks, which are also
distinguished by the NGSIM datasets. It is known that trucks with larger sizes impact on
traffic more greatly than motorcycles and passenger vehicles. However, the NGISM datasets
contain very limited trucks; for example, in the I-80 dataset, more than 80% vehicles are
passenger vehicles with lengths smaller than 7 m (Coifman,2015). Therefore, it is difficult
to explicitly study on trucks’ behavior and impact by using the NGSIM datasets.
In addition, the traffic at intersection are more complicated due to the mixture of pedes-
trians, bicycles (sometimes), and vehicles (Zeng et al.,2014;Lu et al.,2016). Unfortunately,
the NGSIM datasets only contain vehicles, making the studies on the intersection traffic
limited.
•Limitation-5: Relevant data are lacking, such as emissions, demographic and psy-
chological information. Due to the way of collecting the NGSIM trajectories (i.e., cameras),
other data related are still lacking, such as emissions, demographic and psychological infor-
mation. Although simultaneously collecting those data in a large time-space scope is more
like an ”extravagant hope”, there is no doubt that trajectory data with abundant relevant
data will greatly widen and deepen our transportation research (and more papers will appear
in TR-D and TR-F). It may be quite difficult to collect such data under natural conditions,
and controlled experiments may be needed. However, it is obvious that large-scale controlled
8Another reason may be that more high-resolution traffic data at intersection are available (Liu et al.,
2009), because collecting traffic data at intersection is easier that collecting freeway traffic data.
22
experiments will be rather expensive.
7. Outlook
Considering the limitations of the NGSIM datasets, the following outlooks are given.
Note that Outlooks 1-5 are correspondingly given to mitigate Limitations 1-5, and thus the
specific explanations are omitted.
•Outlook-1: High-fidelity trajectory data with larger time-space scopes are expected.
•Outlook-2: High-fidelity trajectory data with various traffic conditions are expected.
•Outlook-3: High-fidelity trajectory data collected on various roads are expected.
•Outlook-4: High-fidelity trajectory data containing various traffic components are
expected.
•Outlook-5: High-fidelity trajectory data with various aspects of related information
are expected.
•Outlook-6: High-fidelity trajectory data with the applications of new technology are
expected. It can be expected that connected vehicles (Uhlemann,2015) and autonomous
vehicles (SAE International,2016) will largely appear on roads in short future, with the
rapidly development of connected and autonomous vehicle technology. Therefore, it will be
meaningful if such mixed traffic data could be collected.
Different from ten years ago, to collect high-fidelity vehicle trajectory data, it may be not
necessary to place cameras in a flying helicopter (Hoogendoorn et al.,2003) or on the top of
a roadside high building (U.S. Federal Highway Administration,2006) any more. Instead, a
rotary wing Unmanned Aerial Vehicle (UAV, or called drone) with tethered power supply is
recommended to carry high definition cameras to collect the high-fidelity traffic data with
a large time-space scope; see Figure 3for an illustration. In addition to the advantages
of the traditional UAV, such as low cost, high flight altitude, and relatively good stability
(Barmpounakis et al.,2016), the tethered UAV is able to stay in sky with much longer time,
such as hours or even days, because the tether cable can sustainably supply power to the
UAV. The advantage of long flight time is significant to collect traffic data with a large time
scope. Nevertheless, UAVs are vulnerable to adverse weather, and advanced technology is
needed to process their collected imagines (Xu et al.,2016a,b). Those disadvantages may
limit their applications in traffic data collection.
8. Summary
This paper reviews the literature conducted based on the well-known high-fidelity NGSIM
trajectory datasets including two freeway-traffic (i.e., US-101 and I-80) datasets and two
arterial-traffic datasets (i.e., Peachtree and Lankershim). Due to the existence of a large
23
Figure 3: A schematic diagram of traffic data collection using a rotary wing UAV with tethered power supply.
amount of relevant literature, only the papers published in the journal series of Transporta-
tion Research are introduced, and all these studies are classified into six subjects, i.e., micro-
scopic traffic flow analysis and modeling,mesoscopic and macroscopic traffic flow modeling,
traffic-related estimation and prediction,traffic flow model calibration,vehicle trajectory data
cleaning, and VANET-related studies.
It is found that the main usages of the NGSIM datasets in the literature are calibrating
or training traffic flow models, validating models, demonstrating driving behaviors or traffic
phenomena, being analysis samples, and building simulation environments or testbeds.
Five limitations (see the highlights in Section 6) of the NGSIM datasets are pointed
out, and six outlooks (see the highlights in Section 7) are accordingly given to shed light
on future traffic flow data collection and data-based studies. Moreover, a rotary wing UAV
with tethered power supply is recommended to carry out the task of traffic data collection in
future, due to its advantages such as low cost, high flight altitude, good stability, and long
flight time.
References
U.S. Federal Highway Administration. Next Generation Simulation Program (NGSIM).
http://ops.fhwa.dot.gov/trafficanalysistools/ngsim.htm 2006.
24
Lijun, S.; Yafeng, Y. Discovering themes and trends in transportation research using topic
modeling. Transportation Research Part C: Emerging Technologies 2017,77, 49–66.
Tordeux, A.; Lassarre, S.; Roussignol, M. An adaptive time gap car-following model. Trans-
portation Research Part B: Methodological 2010,44, 1115–1131.
Koutsopoulos, H.N.; Farah, H. Latent class model for car following behavior. Transportation
Research Part B: Methodological 2012,46, 563–578.
Chen, D.; Laval, J.; Zheng, Z.; Ahn, S. A behavioral car-following model that captures
traffic oscillations. Transportation Research Part B: Methodological 2012,46, 744–761.
Laval, J.A.; Toth, C.S.; Zhou, Y. A parsimonious model for the formation of oscillations in
car-following models. Transportation Research Part B: Methodological 2014,70, 228–238.
Tian, J.; Treiber, M.; Ma, S.; Jia, B.; Zhang, W. Microscopic driving theory with oscilla-
tory congested states: Model and empirical verification. Transportation Research Part B:
Methodological 2015,71, 138–157.
Wang, W.; Zhang, W.; Guo, H.; Bubb, H.; Ikeuchi, K. A safety-based approaching be-
havioural model with various driving characteristics. Transportation Research Part C:
Emerging Technologies 2011,19, 1202–1214.
Przybyla, J.; Taylor, J.; Jupe, J.; Zhou, X. Estimating risk effects of driving distraction:
A dynamic errorable car-following model. Transportation Research Part C: Emerging
Technologies 2015,50, 117–129.
Hamdar, S.H.; Mahmassani, H.S.; Treiber, M. From behavioral psychology to acceleration
modeling: Calibration, validation, and exploration of drivers’ cognitive and safety param-
eters in a risk-taking environment. Transportation Research Part B: Methodological 2015,
78, 32–53.
Zheng, J.; Suzuki, K.; Fujita, M. Car-following behavior with instantaneous driver-vehicle
reaction delay: A neural-network-based methodology. Transportation Research Part C:
Emerging Technologies 2013,36, 339–351.
He, Z.; Zheng, L.; Guan, W. A simple nonparametric car-following model driven by field
data. Transportation Research Part B: Methodological 2015,80, 185–201.
Hao, H.; Ma, W.; Xu, H. A fuzzy logic-based multi-agent car-following model. Transportation
Research Part C: Emerging Technologies 2016,69, 477–496.
25
Chiabaut, N.; Leclercq, L.; Buisson, C. From heterogeneous drivers to macroscopic patterns
in congestion. Transportation Research Part B: Methodological 2010,44, 299–308.
Chen, D.; Laval, J.A.; Ahn, S.; Zheng, Z. Microscopic traffic hysteresis in traffic oscillations:
A behavioral perspective. Transportation Research Part B: Methodological 2012,46, 1440–
1453.
Chen, D.; Ahn, S.; Laval, J.; Zheng, Z. On the periodicity of traffic oscillations and capacity
drop: The role of driver characteristics. Transportation Research Part B: Methodological
2014,59, 117–136.
Wei, D.; Liu, H. Analysis of asymmetric driving behavior using a self-learning approach.
Transportation Research Part B: Methodological 2013,47, 1–14.
Li, L.; Chen, X.; Li, Z. Asymmetric stochastic Tau Theory in car-following. Transportation
Research Part F: Traffic Psychology and Behaviour 2013,18, 21–33.
Taylor, J.; Zhou, X.; Rouphail, N.M.; Porter, R.J. Method for investigating intradriver
heterogeneity using vehicle trajectory data: A Dynamic Time Warping approach. Trans-
portation Research Part B: Methodological 2015,73, 59–80.
Hamdar, S.H.; Qin, L.; Talebpour, A. Weather and road geometry impact on longitudinal
driving behavior: Exploratory analysis using an empirically supported acceleration mod-
eling framework. Transportation Research Part C: Emerging Technologies 2016,67, 193–
213.
Laval, J.A.; Leclercq, L. Microscopic modeling of the relaxation phenomenon using a
macroscopic lane-changing model. Transportation Research Part B: Methodological 2008,
42, 511–522.
Zheng, Z.; Ahn, S.; Chen, D.; Laval, J. The effects of lane-changing on the immediate
follower: Anticipation, relaxation, and change in driver characteristics. Transportation
Research Part C: Emerging Technologies 2013,26, 367–379.
Talebpour, A.; Mahmassani, H.S.; Hamdar, S.H. Modeling lane-changing behavior in a con-
nected environment: A game theory approach. Transportation Research Part C: Emerging
Technologies 2015,59, 216–232.
Balal, E.; Cheu, R.L.; Sarkodie-Gyan, T. A binary decision model for discretionary lane
changing move based on fuzzy inference system. Transportation Research Part C: Emerging
Technologies 2016,67, 47–61.
26
Yang, H.; Jin, W.L. A control theoretic formulation of green driving strategies based on
inter-vehicle communications. Transportation Research Part C: Emerging Technologies
2014,41, 48–60.
Gong, S.; Shen, J.; Du, L. Constrained optimization and distributed computation based car
following control of a connected and autonomous vehicle platoon. Transportation Research
Part B: Methodological 2016,94, 314–334.
Tak, S.; Woo, S.; Yeo, H. Study on the framework of hybrid collision warning system
using loop detectors and vehicle information. Transportation Research Part C: Emerging
Technologies 2016,73, 202–218.
Ahmed, K. Modeling drivers’ acceleration and lane changing behaviors. PhD thesis, MIT,
1999.
Akaike, H. A new look at the statistical model identification. IEEE Transactions on Auto-
matic Control 1974,19, 716–723.
Newell, G. A simplified car-following theory: A lower order model. Transportation Research
Part B: Methodological 2002,36, 195–205.
Kerner, B. Introduction to Modern Traffic Flow Theory and Control: The Long Road to
Three-Phase Traffic Theory; Springer, 2009.
Hamdar, S.H.; Treiber, M.; Mahmassani, H.S. Calibration of a Stochastic Car-Following
Model Using Trajectory Data: Exploration and Model Properties. 88th Annual Meeting
Transportation Research Board; , 2009.
Lee, D. A theory of visual control of braking based on information about time-to-collision.
Perception 1976,5, 437–459.
Kaneman, D.; Tversky, A. Prospect theory: An analysis of decision under risk. Econometrica
1979,47, 263–292.
Chiu, Y.C.; Zhou, L.; Song, H. Development and calibration of the Anisotropic Meso-
scopic Simulation model for uninterrupted flow facilities. Transportation Research Part B:
Methodological 2010,44, 152–174.
Daganzo, C.F. Requiem for second-order fluid approximations of traffic flow. Transportation
Research Part B 1995,29, 277–286.
Piccoli, B.; Han, K.; Friesz, T.L.; Yao, T.; Tang, J. Second-order models and traffic data from
mobile sensors. Transportation Research Part C: Emerging Technologies 2015,52, 32–56.
27
Qian, Z.; Li, J.; Li, X.; Zhang, M.; Wang, H. Modeling heterogeneous traffic flow: A
pragmatic approach. Transportation Research Part B: Methodological 2017,99, 183–204.
Jin, W.L. A kinematic wave theory of lane-changing traffic flow. Transportation Research
Part B: Methodological 2010,44, 1001–1021.
Jin, W.L. A multi-commodity Lighthill-Whitham-Richards model of lane-changing traffic
flow. Transportation Research Part B: Methodological 2013,57, 361–377.
Coifman, B. Empirical flow-density and speed-spacing relationships: Evidence of vehicle
length dependency. Transportation Research Part B: Methodological 2015,78, 54–65.
Siqueira, A.F.; Peixoto, C.J.T.; Wu, C.; Qian, W.L. Effect of stochastic transition in the
fundamental diagram of traffic flow. Transportation Research Part B: Methodological 2016,
87, 1–13.
Jabari, S.E.; Zheng, J.; Liu, H.X. A probabilistic stationary speed-density relation based on
Newell’s simplified car-following model. Transportation Research Part B: Methodological
2014,68, 205–223.
Wu, L.; Coifman, B. Improved vehicle classification from dual-loop detectors in congested
traffic. Transportation Research Part C: Emerging Technologies 2014,46, 222–234.
Laval, J.A. Hysteresis in traffic flow revisited: An improved measurement method. Trans-
portation Research Part B: Methodological 2011,45, 385–391.
Ahn, S.; Vadlamani, S.; Laval, J. A method to account for non-steady state conditions
in measuring traffic hysteresis. Transportation Research Part C: Emerging Technologies
2013,34, 138–147.
Zheng, Z.; Ahn, S.; Chen, D.; Laval, J. Applications of wavelet transform for analysis
of freeway traffic: Bottlenecks, transient traffic, and traffic oscillations. Transportation
Research Part B: Methodological 2011,45, 372–384.
Zheng, Z.; Ahn, S.; Chen, D.; Laval, J. Freeway traffic oscillations: Microscopic analysis of
formations and propagations using Wavelet Transform. Transportation Research Part B:
Methodological 2011,45, 1378–1388.
Blandin, S.; Argote, J.; Bayen, A.M.; Work, D.B. Phase transition model of non-stationary
traffic flow: Definition, properties and solution method. Transportation Research Part B:
Methodological 2013,52, 31–55.
28
Oh, S.; Yeo, H. Impact of stop-and-go waves and lane changes on discharge rate in recovery
flow. Transportation Research Part B: Methodological 2015,77, 88–102.
Li, X.; Cui, J.; An, S.; Parsafard, M. Stop-and-go traffic analysis: Theoretical proper-
ties, environmental impacts and oscillation mitigation. Transportation Research Part B:
Methodological 2014,70, 319–339.
Herrera, J.C.; Bayen, A.M. Incorporation of Lagrangian measurements in freeway traffic
state estimation. Transportation Research Part B: Methodological 2010,44, 460–481.
Deng, W.; Lei, H.; Zhou, X. Traffic state estimation and uncertainty quantification based
on heterogeneous data sources: A three detector approach. Transportation Research Part
B: Methodological 2013,57, 132–157.
Bucknell, C.; Herrera, J.C. A trade-off analysis between penetration rate and sampling
frequency of mobile sensors in traffic state estimation. Transportation Research Part C:
Emerging Technologies 2014,46, 132–150.
Argote-Caba˜nero, J.; Christofa, E.; Skabardonis, A. Connected vehicle penetration rate for
estimation of arterial measures of effectiveness. Transportation Research Part C: Emerging
Technologies 2015,60, 298–312.
Ramezani, M.; Geroliminis, N. On the estimation of arterial route travel time distribution
with Markov chains. Transportation Research Part B: Methodological 2012,46, 1576–1590.
Feng, Y.; Hourdos, J.; Davis, G.A. Probe vehicle based real-time traffic monitoring on urban
roadways. Transportation Research Part C: Emerging Technologies 2014,40, 160–178.
Qi, H.; Wang, D.; Chen, P.; Bie, Y. Simulation of departure flow profile at stop lines for
signal approach spillover. Transportation Research Part C: Emerging Technologies 2013,
33, 88–106.
Srivastava, A.; Jin, W.L.; Lebacque, J.P. A modified Cell Transmission Model with realis-
tic queue discharge features at signalized intersections. Transportation Research Part B:
Methodological 2015,81, 302–315.
Sun, Z.; Ban, X.J. Vehicle trajectory reconstruction for signalized intersections using mobile
traffic sensors. Transportation Research Part C: Emerging Technologies 2013,36, 2680–
283.
Hao, P.; Sun, Z.; Ban, X.J.; Guo, D.; Ji, Q. Vehicle index estimation for signalized intersec-
tions using sample travel times. Transportation Research Part C: Emerging Technologies
2013,36, 513–529.
29
Hao, P.; Ban, X.J.; Guo, D.; Ji, Q. Cycle-by-cycle intersection queue length distribution
estimation using sample travel times. Transportation Research Part B: Methodological
2014,68, 185–204.
Sun, Z.; Zan, B.; Ban, X.J.; Gruteser, M. Privacy protection method for fine-grained urban
traffic modeling using mobile sensors. Transportation Research Part B: Methodological
2013,56, 50–69.
Yang, K.; Guler, S.I.; Menendez, M. Isolated intersection control for various levels of vehicle
technology: Conventional, connected, and automated vehicles. Transportation Research
Part C: Emerging Technologies 2016,72, 109–129.
Lee, S.; Wong, S.C.; Li, Y.C. Real-time estimation of lane-based queue lengths at isolated
signalized junctions. Transportation Research Part C: Emerging Technologies 2015,56, 1–
17.
Lee, S.; Wong, S. Group-based approach to predictive delay model based on incremental
queue accumulations for adaptive traffic control systems. Transportation Research Part
B: Methodological 2017,98, 1–20.
Edie, L. Discussion of Traffic Stream Measurements and Definitions. 2nd International
Symposium on the Theory of Traffic Flow, 1963, pp. 139–154.
Li, X.; Wang, X.; Ouyang, Y. Characterization of traffic oscillation propagation under nonlin-
ear car-following laws. Transportation Research Part B: Methodological 2012,46, 409–423.
Newell, G. A simplified theory of kinematic waves in highway traffic I: General theory. II:
Queuing at freeway bottlenecks. III: Multi-destination flows. Transportation Research Part
B1993,27, 281–313.
Daganzo, C.F. A variational formulation of kinematic waves: Basic theory and complex
boundary conditions. Transportation Research Part B: Methodological 2005,39, 187–196.
Ilgin Guler, S.a.; Menendez, M.a.; Meier, L.a. Using connected vehicle technology to improve
the efficiency of intersections. Transportation Research Part C: Emerging Technologies
2014,46, 121–131.
Li, X.; Wang, X.; Ouyang, Y. Prediction and field validation of traffic oscillation propagation
under nonlinear car-following laws. Transportation Research Part B: Methodological 2012,
46, 409–423.
30
Rhoades, C.; Wang, X.; Ouyang, Y. Calibration of nonlinear car-following laws for traf-
fic oscillation prediction. Transportation Research Part C: Emerging Technologies 2016,
69, 328–342.
Kim, I.; Kim, T.; Sohn, K. Identifying driver heterogeneity in car-following based on a
random coefficient model. Transportation Research Part C: Emerging Technologies 2013,
36, 34–44.
Vieira da Rocha, T.; Leclercq, L.; Montanino, M.; Parzani, C.; Punzo, V.; Ciuffo, B.; Ville-
gas, D. Does traffic-related calibration of car-following models provide accurate estimations
of vehicle emissions? Transportation Research Part D: Transport and Environment 2015,
34, 267–280.
Li, L.; Chen, X.M.; Zhang, L. A global optimization algorithm for trajectory data based
car-following model calibration. Transportation Research Part C: Emerging Technologies
2016,68, 311–332.
Durrani, U.; Lee, C.; Maoh, H. Calibrating the Wiedemann’s vehicle-following model using
mixed vehicle-pair interactions. Transportation Research Part C: Emerging Technologies
2016,67, 227–242.
Zhong, R.X.; Fu, K.Y.; Sumalee, A.; Ngoduy, D.; Lam, W.H.K. A cross-entropy method
and probabilistic sensitivity analysis framework for calibrating microscopic traffic models.
Transportation Research Part C: Emerging Technologies 2016,63, 147–169.
Sopasakis, A.; Katsoulakis, M.A. Information metrics for improved traffic model fidelity
through sensitivity analysis and data assimilation. Transportation Research Part B:
Methodological 2016,86, 1–18.
Chandler, R.E.; Herman, R.; Montroll, E.W. Traffic Dynamics: Studies in Car Following.
Operations Research 1958,6, 165–184.
Aghabayk, K.; Sarvi, M.; Young, W.; Kautzsch, L. A novel methodology for evolutionary
calibration of VISSIM by multi-threading. 36th Australasian Transport Research Forum;
, 2013.
Punzo, V.; Borzacchiello, M.T.; Ciuffo, B. On the assessment of vehicle trajectory data
accuracy and application to the Next Generation SIMulation (NGSIM) program data.
Transportation Research Part C: Emerging Technologies 2011,19, 1243–1262.
Zheng, Z.; Washington, S. On selecting an optimal wavelet for detecting singularities in
traffic and vehicular data. Transportation Research Part C: Emerging Technologies 2012,
25, 18–33.
31
Montanino, M.; Punzo, V. Trajectory data reconstruction and simulation-based valida-
tion against macroscopic traffic patterns. Transportation Research Part B: Methodological
2015,80, 82–106.
Fard, M.R.; Mohaymany, A.S.; Shahri, M. A new methodology for vehicle trajectory recon-
struction based on wavelet analysis. Transportation Research Part C 2017,74, 150–167.
Zheng, Z.; Su, D. Traffic state estimation through compressed sensing and Markov random
field. Transportation Research Part B: Methodological 2016,91, 525–554.
Baiocchi, A. Analysis of timer-based message dissemination protocols for inter-vehicle com-
munications. Transportation Research Part B: Methodological 2016,90, 105–134.
Du, L.; Gong, S.; Wang, L.; Li, X.Y. Information-traffic coupled cell transmission model
for information spreading dynamics over vehicular ad hoc network on road segments.
Transportation Research Part C: Emerging Technologies 2016,73, 30–48.
Treiber, M.; Kesting, A. Validation of traffic flow models with respect to the spatiotem-
poral evolution of congested traffic patterns. Transportation Research Part C: Emerging
Technologies 2012,21, 31–41.
Bifulco, G.N.; Pariota, L.; Brackstione, M.; Mcdonald, M. Driving behaviour models en-
abling the simulation of Advanced Driving Assistance Systems: Revisiting the Action Point
paradigm. Transportation Research Part C: Emerging Technologies 2013,36, 352–366.
Wu, X.; Liu, H.X. Using high-resolution event-based data for traffic modeling and control:
An overview. Transportation Research Part C: Emerging Technologies 2014,42, 28–43.
Jiang, R.; Hu, M.B.; Zhang, H.M.; Gao, Z.Y.; Jia, B.; Wu, Q.S. On some experimental
features of car-following behavior and how to model them. Transportation Research Part
B: Methodological 2015,80, 338–354.
Jin, C.J.; Wang, W.; Jiang, R.; Zhang, H.M.; Wang, H.; Hu, M.B. Understanding the struc-
ture of hyper-congested traffic from empirical and experimental evidences. Transportation
Research Part C: Emerging Technologies 2015,60, 324–338.
Liu, H.X.; Wu, X.; Ma, W.; Hu, H. Real-time queue length estimation for congested signal-
ized intersections. Transportation Research Part C: Emerging Technologies 2009,17, 412–
427.
Guan, W.; He, S. Statistical features of traffic flow on urban freeways. Physica A: Statistical
Mechanics and its Applications 2008,387, 944–954.
32
He, Z.; Zheng, L.; Chen, P.; Guan, W. Mapping to Cells: A Simple Method to Extract Traffic
Dynamics from Probe Vehicle Data. Computer-Aided Civil and Infrastructure Engineering
2017,32, 252–267.
Zeng, W.; Chen, P.; Nakamura, H.; Iryo-Asano, M. Application of social force model to
pedestrian behavior analysis at signalized crosswalk. Transportation Research Part C:
Emerging Technologies 2014,40, 143–159.
Lu, L.; Ren, G.; Wang, W.; Chan, C.Y.; Wang, J. A cellular automaton simulation model
for pedestrian and vehicle interaction behaviors at unsignalized mid-block crosswalks. Ac-
cident Analysis and Prevention 2016,95, 425–437.
Uhlemann, E. Introducing connected vehicles. IEEE Vehicular Technology Magazine 2015,
10.
SAE International. Automated driving levels of driving automation are defined in new SAE
international standard J3016. http://www.sae.org/misc/pdfs/automated driving.pdf 2016.
Hoogendoorn, S.; Van Zuylen, H.; Schreuder, M.; Gorte, B.; Vosselman, G. Microscopic
Traffic Data Collection by Remote Sensing. Transportation Research Record: Journal of
the Transportation Research Board 2003,1855, 121–128.
Barmpounakis, E.N.; Vlahogianni, E.I.; Golias, J.C. Unmanned Aerial Systems for Trans-
portation Engineering: Current practice and future challenges. International Journal of
Transportation Science and Technology 2016,5, 111–122.
Xu, Y.; Yu, G.; Wu, X.; Wang, Y.; Ma, Y. An Enhanced Viola-Jones Vehicle Detec-
tion Method From Unmanned Aerial Vehicles Imagery. IEEE Transactions on Intelligent
Transportation Systems 2016,18, 1845–1856.
Xu, Y.; Yu, G.; Wang, Y.; Wu, X.; Ma, Y. A hybrid vehicle detection method based on
viola-jones and HOG + SVM from UAV images. Sensors 2016,16.
33
