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