Estimating Traffic Flows from Annual
Average Daily Traffic and Third Party
Speed Data
NATMEC
Jaimyoung Kwon and Karl Petty, Iteris
June 30, 2014
2
Objective: Compute Delay
Everybody wants delay and congestion measurements everywhere • We now have speeds everywhere via 3rd party speed (Inrix, HERE, TomTom),
why not flows?
• We just need flows and we’re home free
Many methods • Use existing point detectors and fuse with 3rd party speeds
• Just assume a constant value (eg: 1500 veh/lan/hour)
• Develop flow estimates based solely on probes (can be done by current 3rd party providers)
• Use existing AADT values (flow for a day)
Goal: Estimate hourly vehicular flows for a large number of locations based on AADT
3
Inputs
Annual Average Daily Traffic
(AADT) values coming from
the Federal Highway
Performance Monitoring
System (HPMS)
• And the fact that it’s snapped
to a GIS network (NHPN)
Canonical hourly flow
profiles
Third party speed profiles
4
Flow estimation consists of 4 tasks: 1. Map HPMS links and 3rd party speeds onto common road network
2. Given HPMS AADT for a location, assign daily flow
• Divide AADT by 2 => Assign to each direction
• Adjust by annual growth, monthly, and day of week factors
• End up with a flow per day
3. Generate canonical hourly flow profiles
• Using available flow data from regional sources
• Or just borrow a national profile
4. Pick the most probable hourly flow profile for the given link
• Where the persistent slowdowns take place (eg: AM slowdown => AM peak flow)
• Using third party speed data
5
Conflating PeMS, TMC and HPMS Maps
HPMS Map TMC Map Detector Map
AADT Speeds
PeMS Volumes
Canonical Map
6
Matching Canonical Links with HPMS Link
HPMS link geometry comes from
National Transportation Atlas
Database as ESRI shape files
(updated annually)
AADT values are included as
attributes in the shape files
Map matching between HPMS and
canonical links is assigning AADT
values to canonical links
HPMS links are bidirectional, while
canonical links are directed
HPMS Link
I-5 NB
I-5 SB
Canonical Links
Interstate 5: absolute postmiles 522-524
7
Errors in Map Matching
Errors occur when HPMS arterial
links are too close to Canonical
freeway links
Wrong map matching results in huge
errors in AADT assigned to Canonical
links
Sanity check:
• Opposite canonical link must have the
same AADT value
• Either the nearest upstream or
downstream neighbor must have the
same AADT value
HPMS Link for X St, AADT = 7,539
HPMS Link for I-80, AADT = 126,500
Error: I-80 Canonical Link is matched with X St. HPMS Link
8
Computing Adjustment Factors
𝑻𝒐𝒕𝒂𝒍𝑫𝒂𝒊𝒍𝒚𝑭𝒍𝒐𝒘 =
𝟏
𝟐× 𝑨𝒏𝒏𝒖𝒂𝒍𝑮𝒓𝒐𝒘𝒕𝒉𝑭𝒂𝒄𝒕𝒐𝒓 × 𝑴𝒐𝒏𝒕𝒉𝑶𝒇𝒀𝒆𝒂𝒓𝑭𝒂𝒄𝒕𝒐𝒓 × 𝑫𝒂𝒚𝑶𝒇𝑾𝒆𝒆𝒌𝑭𝒂𝒄𝒕𝒐𝒓 × 𝑨𝑨𝑫𝑻
Leverage regional data – looking for general trends
• Use vehicle counts for an average day of the past 12 months to obtain annual growth factor
• Use vehicle counts for an average day of every month to compute MoY adjustment factor
• Use vehicle counts for an average Mon-Sun to compute DoW adjustment factor
Each of these adjustment factors can be tagged by a geographic region – U.S. (default),
state, county
Each of these adjustment factors can be tuned for a particular daily volume range
9
Month of Year Factors
Caltrans District 3 VDS 314013:
average weekly volumes over 12
months
Estimated MOY adjustment factor
using data from 415 detectors in
Caltrans District 3
0 2 4 6 8 10 123
3.5
4
4.5x 10
5
Month
Ave
rage
Wee
kly
Volu
me
0 2 4 6 8 10 120.7
0.8
0.9
1
1.1
1.2
1.3
1.4
1.5
Month
Mon
th o
f Yea
r Fac
tor
10
Day of Week Factors
Caltrans District 3 VDS 317213:
Day of Week flows
Estimated DOW adjustment factor
using data from 415 detectors in
Caltrans District 3
1 2 3 4 5 6 70.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Day of Week
Day
of W
eek
Fact
or
11
Computing Hourly Flow Profiles
Need a small number of canonical
flow profiles
We’re going to assign one to each
location
• Function of day of week
We decided on 4 basic patterns:
AM peak, PM peak, midday peak
and double peak
Profiles can be also made by
geographic region, daily volume
range and season
For each profile, find the mean and
normalize
AM peak PM peak
Double peak Midday peak
12
Grouping Hourly Flow Distributions
Flow profiles of the same type (e.g. PM peak) are different for different daily volumes
We did not find it useful maintaining different flow profile of the same type for different
seasons or days of week (we still have a day of week factor, though)
0 5 10 15 20 250
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Time (hours)
PM peak for links with daily volume <20K
0 5 10 15 20 250
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
Time (hours)
PM peak for links with daily volume >100K
13
Using Speed to Determine Flow Distribution
Determine free flow speed by looking at speed profile between 5am and 11pm
Find time intervals where average hourly speed drops below 90% of free flow: before
10 am = AM peak, after 3 pm = PM peak, 10 am – 3 pm = midday peak
Presence of both AM and PM peaks = double peak
Double peak AM peak PM peak
14
Challenges of Flow Classification
Sometimes, it is impossible to infer
anything from speed
measurement
Sometimes, speed drop occurs
due to accident and triggers wrong
flow distribution choice
AM peak guessed
Ground truth indicates PM peak
Speed drop due to incident
No significant speed drop
15
Refinements of Speed Classification
If nothing can be inferred from speed, use default flow distribution
• Double peak for workdays
• Midday peak for weekends and holidays
For those links with default flow distribution, look at upstream and downstream
neighbors and borrow from them
If upstream and downstream neighbors have different flow distribution, look at
the opposite link
• AM peak on the opposite link implies PM peak on the current link
• PM peak on the opposite link implies AM peak on the current link
• Otherwise, use the same flow distribution
Use speed data for multiple days to obtain a speed profile for a given Day of
Week; and then work with that speed profile to determine flow distribution – this
reduces the impact of incidents
16
Using Caltrans D3 Data for Evaluation
Used Caltrans District 3 vehicle
counts to
• Determine annual growth factor
(together with HPMS AADT)
• Determine Month of Year and Day of
Week adjustment factors
• Create flow distributions
• Evaluate the results
Used 3rd party speed data for
March 2014 to select flow
distributions
Built prototype that estimates
flows on Caltrans District 3
freeways for March 2014
AM peak
PM peak
Double peak
Midday peak
Vehicle Detector
Default
AADT
17
Example of Flow Estimation
I-5 SB
I-5 NB
18
0
5
10
15
20
25
30
35
40
1 2 3 4
Some Results on Flow Estimation Quality 415 vehicle detector locations
Random date: Monday, March 24, 2014
4 experiments, each improving on the
previous one: 1. Initial try – simple map matching between HPMS and
canonical links; links without speed data are
considered to be without flow
2. Use default values and flow distributions for
canonical links without data
3. Smear properly estimated flows to the nearest
neighboring links without data
4. Use mapping of HPMS links onto canonical links with
sanity check; and refine speed classification for
picking flow distribution
Metrics • Number of correct flow distribution choices out of 415
• Mean percentage error of hourly flows over 415 links
with detectors
0
50
100
150
200
250
300
350
1 2 3 4
Experiment #
Experiment #
Hourly Flow % error
Number of Correct Flow Distribution Choices
19
Computing Performance Measures
3.78 mile stretch of Interstate 5 South in
Sacramento: from postmile 525.78 to
postmile 522
Compute hourly Vehicle Miles Traveled
(VMT), Vehicle Hours Traveled (VHT) and
Delay (vehicle-hours) in 3 ways
1. Spot detectors (flows and speeds)
2. Using spot flows and 3rd party speed data
3. Using estimated flows and 3rd party speed
data
Delay is computed for free flow speed 50
mph
Date: Monday, March 24, 2014
20
Computing Performance Measures (cont.)
0 5 10 15 20 250
1
2
3x 10
4
VM
T
0 5 10 15 20 250
200
400
600
VH
T
0 5 10 15 20 250
20
40
Dela
y (
Veh
.-H
ou
rs)
Time (Hours)
PeMS
PeMS flows + HERE Speeds
Estimated flows + HERE speeds
Estimate of delay
for a single location
Extending to many
locations now
21
Summary
(still a work in progress)
Using AADTs and 3rd party speeds to estimate flow seems
suitable for computing traffic performance measures
• Might be useful for setting static ramp metering plans
• It cannot replace traffic volume measurements for feedback traffic
control