Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Vincenzo Punzo
University of Naples Federico II, Italy
Traffic Flow Theory and Characteristics Committee (AHB45)
2014 Summer Meeting - August 11-13, 2014 - Portland, Oregon USA
Symposium Celebrating 50 Years of Traffic Flow Theory
Exploring the impact of microscopic features of traffic
on macroscopic patterns
Contributors
• Marcello Montanino
• Biagio Ciuffo (now at the European Commission Joint Research
Centre)
The twofold representation of traffic provided by NGSIM(-like) data
• NGSIM: trajectory of each single vehicle crossing the observed time-
space domain spatiotemporal traffic patterns from microscopic data
• Other traffic data: very partial view of the traffic phenomenon
Time-space trajectory plot
sp
ace
Time
Time-space Edie’s speed contour plot
Time
sp
ace
• The traffic scientific community has not yet thoroughly exploited
this twofold ‘value’ of NGSIM data
Twofold representation and TFT studies (1)
In microscopic modelling of traffic:
• We rather know how each single component works alone:
Theorethical investigations and empirical analyses of single ‘driver’ models
(CF or LC models)
In car-following, most often single-lane, homogeneous-flow assumptions
Calibration and validation on disaggregate data focused on riproducing the
single behaviour
• We do not really know what happens when single components
interact in a stochastic traffic simulation environment:
Interaction effect of LC and CF decisions (i.e. of models output)?
Impact of parameter heterogeneity?
Impact of assumptions on the probabilistic model of parametric inputs?
• What is the impact of microscopic features and assumptions on the
collective behaviour of traffic? i.e. as captured by spatiotemporal
congested patterns?
Spatiotemporal patterns are not the most often used measure to
investigate microscopic models (often in simple settings like e.g. car-
following only; Treiber & Kesting, 2012, TR-C)
• Dichotomy between microscopic TFT and traffic micro-simulation
Traffic simulation outcomes mostly investigated by means of time series
or frequency plots
Twofold representation and TFT studies (2)
Methodology in extreme synthesis
• Using measured and simulated macroscopic
spatiotemporal traffic patterns to investigate impact
of microscopic features
• Calibrating on disaggregate data and validating on
collective data
Analysis framework
Analysis of macroscopic traffic patterns
Raw NGSIM trajectories
Time [s]
Spac
e[m
]
Reconstructed traject.
Time [s]
Spac
e[m
]
Estimated modelparameters
for each vehicle
Disaggregate Calibration
Trace-driven Simulation
Aggregation of measurements
Filtering & Reconstr.
CF LC
Speed
[kph]Speed
[kph]
Impact of microscopic
features like:
• Modelling assumptions
• Calibration methods and
settings
• Set of parameters to
calibrate
• Assumptions on
parameters’ pdf
• …
Calibration of IDM and MOBIL (Treiber et al., 2000, Ph. Rev. E; Kesting et al., 2007, TRR)
• IDM:
min𝛽
RMSE 𝑠𝑝𝑎𝑐𝑖𝑛𝑔
OptQuest Multistart (Punzo et al. 2012, TRR)
Uncertainty analysis and physical informed criteria to set parameters
bounds (Punzo et al. 2014, IEEE T-ITS)
• MOBIL (no calibration attempts in literature; (*) Zheng, 2014, TR-B):
The lane-changing event is rare not interested in the instant and its
traffic conditions but in the prevailing traffic conditions that generate it
scenario leaders & followers in current and target lanes do not change
min𝛽
#LC
𝜎𝐿𝐶+
#noLC
𝜎𝑛𝑜𝐿𝐶; where:
#LC = number of unsuccessful LC scenarios (1-detection rate(*))
#noLC = number of unsuccessful noLC scenarios (false alarm rate(*))
𝜎𝐿𝐶 e 𝜎𝑛𝑜𝐿𝐶 computed through Monte Carlo uncertainty analysis
Calibration of MOBIL (2)
MOBIL LC model calculates the potential advantage of all the
vehicles involved in the LC maneuver in terms of acceleration, as
given by the IDM model;
MOBIL calibration is therefore conditional on the IDM calibration;
In calibration:
vehicles are moved according to the NGSIM measurements
(and not using the IDM).
The IDM is used to calculate accelerations yielding the potential
advantage, by using for each vehicle its own IDM calibrated
parameters;
Trace driven micro-simulation (NGSIM I80)
• Trace-driven traffic micro-simulation fair comparison
• Vehicle insertion as in measured data
• Downstream conditions superimposed to exiting vehicles
• …
• Car-following: IDM
• Lane Changing: MOBIL
• Merging: mandatory lane-changing with MOBIL
NGSIM DATA ERROR ANALYSIS AND RECONSTRUCTION
Background on analysis of NGSIM data
Thieman, Treiber & Kesting, 2008, TRR
• Because of the noise in the positional data, velocity and
acceleration information cannot be extracted directly
• Symmetric exponential moving average filter to be applied “to
all trajectories before any further data analysis”
Punzo, Borzacchiello & Ciuffo, 2009, TRB; 2011, TR-C
• General methodology to quantify the degree of accuracy/bias in
vehicle trajectory data
• Different criteria:
Analysis of accelerations, jerks, amplitude frequency spectrum
Internal consistency, platoon consistency
• Application to all the NGSIM datasets
• High level of measurement errors
• 4.0-12.4% of leader-follower couples show unphysical inter-
vehicle spacing
Opened issues
• To which analyses can NGSIM data be reliably applied to without
any processing?
• Which is the impact of such errors on analyses made on NGSIM
data?
• Which the accuracy requirements for future data gathering?
• …
• Since 2012, 5 out of 19 studies published on journals(*) using
NGSIM trajectories applied some kind of filter to data:
2 out of 11 on car-following
3 out of 8 on lane-changing
(*) source Scopus, accessed 28/07/2014
Filtering NGSIM data
• Usual filtering techniques are inadequate (e.g. kernel smoothing)
• If you cannot filter the data, just throw them away...
• If the time window of eliminated points is short, lots of available
information/constraints to reconstruct the missing trajectory:
space travelled within the window;
physical capabilities of cars;
car-following dynamics (inter-vehicle spacing integrity).
Multi-step reconstruction procedure
Montanino and Punzo 2011, TRR
Montanino and Punzo 2015, TRB
Application to I80-1 dataset (available at www.multitude-project.eu )
Inadequateness of low-pass filters
0 10 20 30 40 50 600
5
10
15
20
25
30
35
40
45
50
Time [s]
Sp
eed
[kp
h]
Comparison of filtering techniques
NGSIM Data
Low-pass (0.75 Hz)
Low-pass (0.25 Hz)
Proposed Technique
0 10 20 30 40 50 60-30
-20
-10
0
10
20
30
Time [s]
Accele
rati
on
[m
/s2]
Acceleration Profile
NGSIM Data
Low-pass (0.75 Hz)
Low-pass (0.25 Hz)
Proposed Technique
Removal of: 1) outliers and 2) noise
0 10 20 30 40 50 600
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Speed Profile after step 1
NGSIM data
Filter Step 1
0 10 20 30 40 50 60-30
-20
-10
0
10
20
30
Time [s]
Acc
ele
rati
on
[m
/s2]
Acceleration Profile after step 1
NGSIM data
Filter Step 1
0 10 20 30 40 50 600
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Speed Profile after step 2
Filter Step 1
Filter Steps 1+2
0 10 20 30 40 50 60-30
-20
-10
0
10
20
30
Time [s]
Ac
ce
lera
tio
n [
m/s
2]
Acceleration Profile after step 2
Filter Step 1
Filter Steps 1+2
3) reconstruction 4) residual noise removal
0 10 20 30 40 50 600
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Speed Profile after step 3
Filter Steps 1+2
Filter Steps 1+2+3
0 10 20 30 40 50 60-30
-20
-10
0
10
20
30
Time [s]
Ac
ce
lera
tio
n [
m/s
2]
Acceleration Profile after step 3
Filter Steps 1+2
Filter Steps 1+2+3
0 10 20 30 40 50 600
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Speed Profile after step 4
Filter Steps 1+2+3
Filter Steps 1+2+3+4
0 10 20 30 40 50 60-30
-20
-10
0
10
20
30
Time [s]
Acc
ele
rati
on
[m
/s2]
Acceleration Profile after step 4
Filter Steps 1+2+3
Filter Steps 1+2+3+4
Imposing platoon consistency
8600 8700 8800 8900 9000 910050
100
150
200
250
300
350
400
450
3095
3095
3108
3108
Frame
Sp
ac
e [
m]
Example of circular dependency between two vehicles
Lane 3
Lane 4
8900 9000 9100 9200 9300 9400 9500 9600 9700 9800
100
150
200
250
300
350
400
Frame
Sp
ac
e [
m]
Complex circular dependency among 53 vehicles
Lane 1
Lane 2
Lane 3
Lane 4
Lane 5
Lane 6
Lane 7
8900 9000 9100 9200 9300 9400 9500 9600 9700 9800
100
150
200
250
300
350
400
Frame
Sp
ac
e [
m]
Complex circular dependency among 53 vehicles
Lane 1
Lane 2
Lane 3
Lane 4
Lane 5
Lane 6
Lane 7
8900 9000 9100 9200 9300 9400 9500 9600 9700 9800
100
150
200
250
300
350
400
Frame
Sp
ac
e [
m]
Complex circular dependency among 53 vehicles
Lane 1
Lane 2
Lane 3
Lane 4
Lane 5
Lane 6
Lane 7
8970 9030
8900 9000 9100 9200 9300 9400 9500 9600 9700 9800
100
150
200
250
300
350
400
Frame
Sp
ac
e [
m]
Complex circular dependency among 53 vehicles
Lane 1
Lane 2
Lane 3
Lane 4
Lane 5
Lane 6
Lane 7
Raw vs. Reconstructed trajectories (I80-1)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
0.01
0.02
0.03
0.04
0.05
0.06
Frequency [Hz]
Am
plitu
de
Frequency Spectra of Accelerations from Dataset: I80-1
Raw
Reconstructed
<-10 -8 -6 -4 -2 0 2 4 6 8 >100
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Acceleration [m/s2]
Fre
qu
en
cy
Distribution of Accelerations from Dataset: I80-1
Raw
Reconstructed
7% < -5 m/s2
13% > 3 m/s2
0 5 10 15 20 25 >300
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Maximum Speed [m/s]
Fre
qu
en
cy
Distribution of Maximum Speeds of each vehicle from Dataset: I80-1
Raw
Reconstructed
4% > 110 kph
<-10 0 10 20 30 40 >500
0.05
0.1
0.15
0.2
0.25
Minimum Spacing [m]
Fre
qu
en
cy
Distribution of Minimum Spacings from Dataset: I80-1
Raw
Reconstructed
7% < 0 m
Raw vs. reconstructed macroscopic patterns resolution 10m x 10s
Aggregating trajectory data, differences are barely visible
1
Speed
[kph]
Speed
[kph]
IMPACT OF MEASUREMENT ERRORS ON CALIBRATION OF DRIVER MODELS (CF AND LC)
Impacts of trajectory measurement errors: background
• Ossen and Hoogendoorn, 2008, TRR
Impact of errors on calibration of car-following models
Experiment with synthetic data
white Gaussian noise assumption
Significant impact of errors on calibration results
• Only impact on calibration: no lane changing and no impact on
aggregate results that is on traffic simulation outputs
Analysis of macroscopic traffic patterns
Reconstructed traject.
Time [s]
Spac
e[m
]
Estimated modelparameters
for each vehicle
Disaggregate Calibration
Trace-driven Simulation
Aggregation of measurements
CF LC
Speed
[kph]Speed
[kph]
Raw NGSIM traject.
Time [s]Sp
ace
[m]
Estimated model parameters
for each vehicle
Disaggregate Calibration
LC CF
Trace-driven Simulation
Speed
[kph]
Impact of measurement errors on simulation
Frequencies of calibrated IDM parameters
1 2
3
2 4 6 8 100
0.1
0.2
0.3
0.4
0.5
Fre
qu
en
cy
Frequency plot of
Calibration against Raw data
Calibration against Reconstructed data
0.5 1 1.5 2 2.5 3 3.5 40
0.1
0.2
0.3
0.4
0.5
amax
[m/s2]
Fre
qu
en
cy
Frequency plot of amax
Calibration against Raw data
Calibration against Reconstructed data
22 24 26 28 300
0.1
0.2
0.3
0.4
0.5
Vmax
[m/s]
Fre
qu
en
cy
Frequency plot of Vmax
Calibration against Raw data
Calibration against Reconstructed data
0.5 1 1.5 2 2.5 30
0.02
0.04
0.06
0.08
0.1
0.12
0.14
T [s]
Fre
qu
en
cy
Frequency plot of T
Calibration against Raw data
Calibration against Reconstructed data
0.5 1 1.5 2 2.50
0.1
0.2
0.3
0.4
0.5
bnorm
[m/s2]
Fre
qu
en
cy
Frequency plot of bnorm
Calibration against Raw data
Calibration against Reconstructed data
0.5 1 1.5 2 2.5 30
0.1
0.2
0.3
0.4
0.5
S0 [m]
Fre
qu
en
cy
Frequency plot of S0
Calibration against Raw data
Calibration against Reconstructed data
H0(different distributions) rejected (5% level of significance)
The car-following model acts as a filter
1
2
3
0 10 20 30 40 50 60 700
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Simulated Speed
after calibration against Raw data
Measured Leader
Measured Follower
Simulated Follower
0 10 20 30 40 50 60 700
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Simulated Speed
after calibration against Reconstructed data
Measured Leader
Measured Follower
Simulated Follower
0 10 20 30 40 50 60 700
10
20
30
40
50
Time [s]
Sp
ee
d [
kp
h]
Simulated Speed
Simulated Follower (calibration against Raw data)
Simulated Follower (calibration against Reconstructed data)
a) b)
c)
Negligible impact of
errors on calibration
results.
Ossen and
Hoogendoorn, 2008?
1 2
3
4
<-1 -0,5 0 0,5 >10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Residual (each 0.1 s) [m]
Fre
qu
en
cy
Frequency plot of Residuals on Positions
Estimated Normal Residuals
Observed Residual (mean=0.035 m, std=0.520 m)
-10 0 10 20 30 40 50 60
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99 0.9970.999
Residual (each 0.1 s) [m]
Pro
bab
ilit
y
Normal Probability Plot
Normal
data2
-10 0 10 20 30 40 50 60
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99 0.9970.999
Residual on LocalY [m]
Pro
bab
ilit
y
Normal Probability Plot
Normal
data2
Empirical
-10 0 10 20 30 40 50 60
0.0010.0030.01 0.02 0.05 0.10
0.25
0.50
0.75
0.90 0.95 0.98 0.99 0.9970.999
Residual on LocalY [m]
Pro
bab
ilit
y
Normal Probability Plot
Normal
data2
Empirical
0 10 20 30 40 50 60-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Lag [s]
Part
ial A
uto
co
rrela
tio
n
Partial Autocorrelation Function
95% confidence bounds
0 10 20 30 40 50 60-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Lag [s]
Partial Auto
correlation
Partial Autocorrelation Function
95% confidence bounds
0 10 20 30 40 50 60-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Lag [s]
Part
ial A
uto
co
rrela
tio
n
Partial Autocorrelation Function
95% confidence bounds
0 10 20 30 40 50 60-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Lag [s]
Part
ial A
uto
co
rrela
tio
n
Partial Autocorrelation Function
95% confidence bounds
0 10 20 30 40 50 60-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Lag [s]
Partial Auto
correlation
Partial Autocorrelation Function
95% confidence boundsc)
a) b)
Hp: Measurement errors = residuals between raw and reconstructed positions
Residuals are
not normal
and are auto-
correlated
Frequencies of calibrated MOBIL parameters
1 1 2 3 4 5 6 7 8
0
0.1
0.2
0.3
0.4
0.5
bsafe
[m/s2]
Fre
qu
en
cy
Frequency plot of bsafe
Calibration against Raw data
Calibration against Reconstructed data
0 0.5 1 1.5 20
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
aTh
[m/s2]
Fre
qu
en
cy
Frequency plot of aTh
Calibration against Raw data
Calibration against Reconstructed data
0 0.2 0.4 0.6 0.8 10
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
p
Fre
qu
en
cy
Frequency plot of p
Calibration against Raw data
Calibration against Reconstructed data
H0(different distributions) not rejected (5% level of significance)
Correlation matrices of «raw and clean parameters» (i.e. parameters calibrated against raw and
reconstructed data)
Correlation Matrix - Calibration against Raw data
alphaminTimeHeadmaxV maxAccnormDec s0estSafeFollDecpminThresholdLC
alpha
minTimeHead
maxV
maxAcc
normDec
s0
estSafeFollDec
p
minThresholdLC
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Correlation Matrix
Calibration against Raw data
α
α
T
Vmax
amax
bnorm
ΔS0
bsafe
p
ΔaTh
T Vmax amax bnorm ΔS0 bsafe p ΔaTh
Analysis of parameters correlation (Kim and Mahmassani, 2012, TRR)
Correlation Matrix - Calibration against Reconstruted data
alphaminTimeHeadmaxV maxAccnormDec s0estSafeFollDecpminThresholdLC
alpha
minTimeHead
maxV
maxAcc
normDec
s0
estSafeFollDec
p
minThresholdLC
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Correlation Matrix
Calibration against Reconstructed data
α
α
T
Vmax
amax
bnorm
ΔS0
bsafe
p
ΔaTh
T Vmax amax bnorm ΔS0 bsafe p ΔaTh
• Correlation among CF parameters does not change;
• correlation of amax CF parameter with LC parameters increases (in abs.);
• correlation among LC parameters increases.
• Hp. explanatory capability of the LC model is reduced by errors
Summary of impacts on calibration
• Not normal, auto-correlated residuals
• No impact on car-following calibration results (both on
parameters PDF and correlation structure)
• Impact on Lane-changing calibration results (both on parameters
PDF and correlation structure)
• …
• Analysis results not informative on the impacts on traffic
simulation
Remarks
No car-following calibration possible on raw data without ‘tricks’
(negative inter-vehicle spacing)
To calibrate a CF and LC model over a whole set of 2000
trajectories is not just running 2000 times an optimisation
algorithm!
Major critical steps:
• To verify the calibration setting (MoP, GoF, algorithm)
• To set the bounds for the parameters (Punzo et al. 2014, IEEE T-ITS)
Uncertainty analysis
Physical criteria
Verification after simulation
IMPACT OF MEASUREMENT ERRORS ON SIMULATION OUTPUTS AND MACROSCOPIC TRAFFIC PATTERNS
Analysis of macroscopic traffic patterns
Reconstructed traject.
Time [s]
Spac
e[m
]
Estimated modelparameters
for each vehicle
Disaggregate Calibration
Trace-driven Simulation
Aggregation of measurements
CF LC
Speed
[kph]Speed
[kph]
Raw NGSIM traject.
Time [s]Sp
ace
[m]
Estimated model parameters
for each vehicle
Disaggregate Calibration
LC CF
Trace-driven Simulation
Speed
[kph]
Impact of measurement errors on simulation
«Raw vs. clean parameters» simulations
0 20 40 60 80 100 1200
0.02
0.04
0.06
0.08
0.1
0.12
0.14
Travel Time [s]
Fre
qu
en
cy
Frequency plot of Travel Times
Measured
Simulated with parameterscalibrated againstRaw data
Simulated with parameterscalibrated againstReconstructed data
1000 2000 3000 4000 5000 6000 7000 8000 90000
50
100
150
200
250
300
350
Frame
De
ns
ity
[v
eh
/km
]
Section Density
Real Raw
Measured
Simulated with parameters calibrated against Raw data
Simulated with parameters calibrated against Reconstructed data
1000 2000 3000 4000 5000 6000 7000 8000 90000
100
200
300
400
500
Frame
Den
sit
y [
ve
h/k
m]
Section Density
Real Raw
Measured
Simulated with parameters calibrated against Raw data
Simulated with parameters calibrated against Reconstructed data
1000 2000 3000 4000 5000 6000 7000 8000 90000
100
200
300
400
500
Frame
Den
sit
y [
ve
h/k
m]
Section Density
Real Raw
Measured
Simulated with parameters calibrated against Raw data
Simulated with parameters calibrated against Reconstructed data
1 2 3 4 5 6 On-Ramp0
50
100
150
200
250
300
350
Lane ID
Number of Lane-Changes Per Lane
Nu
mb
er
of
La
ne
-Ch
an
ge
s MeasuredSimulated with parameters calibratedagainst Raw data
Simulated with parameters calibratedagainst Reconstructed data
Speed
[kph]Speed
[kph]
Real measurements
«Clean parameters» sim. «Raw parameters» sim.
Measured vs. simulated Edie’s speed
Remarks
• Impacts of errors on simulation less than expected, given the big
errors in raw data
• Measures other than macroscopic patterns could not capture the
true behavior of models
• In both the simulations congestion patterns are different from that
measured:
congestion is lower at the beginning of the stretch, higher at the end.
Models are not able to reproduce the full upstream propagation of
congestion
• Yet, the ‘clean data simulation’ provided slightly better description
of traffic at both disaggregate and collective description
IMPACTS OF ASSUMPTIONS ON THE INPUT PARAMETER PROBABILITY DENSITY FUNCTIONS (PDF)
Assigning calibrated parameters to vehicles
Reconstructed traject.
Time [s]
Spac
e[m
]
Estimated modelparameters
for each vehicle
Disaggregate Calibration
CF LC
Sampling fromEmpirical Joint
Distribution
Sampling fromEmpirical Marginal
Distributions
Sampling fromNormal Marginal
Distributions
AssigningVehicle-specific calibrated
parameters
Deterministic Sim.
Stochastic Simulations Stochastic Simulations Stochastic Simulations
Sampling fromUniform Marginal
Distributions
Stochastic Simulations
Real measurements
Uniform marginal Normal marginal Empirical marginal
Veh-specific param. Empirical joint
Real Measurements vs. best replication of scenarios
Virtual queue at the entrance
Empirical marginal
106
107
0
0.2
0.4
0.6
0.8
1Empirical Cumulative Distribution Function of the SSEs
SSE [m2/s
2]
Cu
mu
lati
ve
Fre
qu
en
cy
Vehicle-specific CALIBRATED par.
Sampling from EMPIRICAL JOINT
Sampling from EMPIRICAL marginal
Sampling from NORMAL marginal
Sampling from UNIFORM marginal
Empirical CDF of scenario replications’ SSEs
105.56
105.58
105.6
105.62
105.64
105.66
105.68
0
0.2
0.4
0.6
0.8
1Empirical Cumulative Distribution Function of the SSEs
SSE [m2/s
2]
SSE = Sumij (euclidean normij)
In general, non robust for pattern recognition
but suitable for ‘constrained’ patterns like these
Remarks
• Huge impact on results of the assumption on the input
parameter PDFs
• CDFs on SSE offer a clear ranking of the performances of
different input PDFs
• Sampling from ‘empirical marginal’ PDF (i.e. no correlation)
yields ‘unrealistic’ parameters combinations (i.e. virtual queues
at the entrance)
• Sampling from normal marginal, all the behaviours are averaged
no congestion propagation
• If the parameter correlation structure is unknown, the safest
assumption is to sample the parameters from uniform PDFs
(customary assumption in case of no prior information on the
PDF)
Summary (1)
• Impact of measurement errors in trajectory data substantially
neglected in the field literature.
• Enhanced methodology to reconstruct trajectory data,
accounting for inter-vehicle spacing consistency;
• Application to the NGSIM I80-1 dataset;
• Straightforward methodology to evaluate the impacts of
microscopic features on collective behaviour of traffic based on:
consistent calibration of individual driver models (lane changing and
car-following) over the entire trajectory dataset
trace-driven traffic simulation over the same time-space domain of
trajectory data.
New methodology to calibrate rule-based lane changing models
(e.g. MOBIL), based on the concept of lane-changing scenario.
Quantification of measurement errors impact on both car-
following and lane changing calibration:
Negligible impact on CF
Quantified impact on LC
Summary (2)
• Quantification of measurement errors impact on the collective
behaviour of traffic;
• Analysis of the impact of assumptions on the probabilistic model
of inputs on the simulation results, in terms of macroscopic
spatiotemporal traffic patterns.
Conclusions
• Previous studies using NGSIM-like data rarely made use of
collective description of traffic, as resulting from macroscopic
spatiotemporal traffic patterns, to corroborate models and model
assumptions
• NGSIM-like data open up new horizons in researching traffic flow
theory and simulation, enabling the study of the collective
behavior of traffic resulting from single driver models (i.e. car-
following and lane-changing)
• Investigations made on NGSIM-like data will hopefully contribute
to solve the dichotomy between TFT and traffic simulations as
well as micro/macro dualism
• New data gathering efforts of NGSIM-like data are needed
around the world
Main work references
Montanino M. and Punzo V. 2015. Errors in vehicle trajectory data and their impact on
simulated spatiotemporal traffic patterns. Submitted for presentation to the 2015 TRB
Annual meeting, Washington D.C.
Punzo V., M. Montanino and Ciuffo B., 2014. Do we really need to calibrate all the
parameters? Variance-based sensitivity analysis to simplify microscopic traffic flow
models. IEEE Transactions on Intelligent Transportation Systems (in press)
M. Montanino, V. Punzo, 2013. Making NGSIM data usable for studies on traffic flow
theory: a multistep method for vehicle trajectory reconstruction. Transportation
Research Record 2390, p.99-111, TRB.
Punzo V., Ciuffo B. and Montanino M. 2012. Can results of car-following model
calibration based on trajectory data be trusted? Transportation Research Record 2315,
p. 11-24. “Greenshields prize 2012”
Punzo V., Borzacchiello M., Ciuffo B., 2011. On the assessment of vehicle trajectory
data accuracy and application to the Next Generation SIMulation (NGSIM) program
data. Transportation Research Part C 19 (2011), 1243–1262.
www.multitude-project.eu
Contacts:
Vincenzo Punzo
University of Naples Federico II, Italy