Subcycle-based Neural Network Algorithms for Turning Movement

Count Prediction

Yashaswi Karnati

, Rahul Sengupta

, Anand Rangarajan

and Sanjay Ranka

Department of Computer and Information Science & Engineering, University of Florida, Gainesville, FL, U.S.A.

Keywords:

Deep Learning, Intelligent Transportation, Machine Learning, Neural Networks, Turning Movement Count

Prediction.

Abstract:

Predicting intersection turning movements is an important task for urban trafﬁc analysis, planning, and signal

control. However, trafﬁc ﬂow dynamics in the vicinity of urban arterial intersections is a complex and nonlin-

ear phenomenon inﬂuenced by factors such as signal timing plan, road geometry, driver behaviors, queuing,

etc. Most current methods focus on predicting turning movement counts using data at coarser aggregations

in the order of minutes or above. Important details such as platoon movements may be lost at such coarse

resolutions. In this work, we propose machine learning approaches to imputing turning movement counts at

intersections using data at subcycle resolutions, from 5 seconds to 375 seconds. In particular, we show that

deep neural networks are capable of directly learning an abstract representation of intersection trafﬁc dynam-

ics using detector actuation waveforms and signal state information. We generate a large dataset of 30 million

cycles by approximately replicating real-world trafﬁc arrival patterns from archived loop detector data in a

microscopic trafﬁc simulator. We extensively evaluate our models and show that our models predict turning

movement counts with greater accuracy when higher resolution data are provided.

1 INTRODUCTION

Turn movement counts (TMCs) are used for a wide

variety of applications related to intersection analy-

ses, intersection design, and transport planning. Re-

cently, several advances in trafﬁc control technolo-

gies and applications are driving additional needs for

continuous, real-time, quality TMC data. These tech-

nologies include:

1. Adaptive control technologies: These systems dy-

namically change the signal phasing pattern lo-

cally at an intersection based on trafﬁc demand.

2. Regional Integrated Corridor Management Sys-

tem (R-ICMS): This system will perform a meso-

scopic simulation of the network using TMCs to

validate the improvement to the network if a di-

version route response plan is implemented or an

optimized set of signal timing plans is deployed.

These applications have introduced additional

trafﬁc data needs beyond what has been provided with

https://orcid.org/0000-0002-2512-1250

https://orcid.org/0000-0001-9793-5176

https://orcid.org/0000-0001-8695-8436

https://orcid.org/0000-0003-4886-1988

the traditional TMC data collection efforts, as fol-

lows:

1. Temporal coverage: The data need to be provided

for intersections continuously, not just for a sam-

pling of time for a few hours or days.

2. Spatial coverage: The data need to be provided

for intersections throughout an entire corridor or

network in order to support a regional application.

3. Cost effective: The data need to be able to be col-

lected in a cost-effective manner, which requires

the use of technology instead of manual counts.

This derives from the need for temporal and spa-

tial coverage.

4. Real-time: The data need to be available in real-

time for use by real-time operations.

5. Higher granularity: The TMC data need to be ac-

curate for several cycles to a few hours.

The traditional method of using a human to ob-

tain this information manually is very time consuming

and monotonous. Although approach volumes can be

easily derived from the loop detectors, they only give

information on how many vehicles are entering the in-

tersection and not the TMCs. The focus of this work

Karnati, Y., Sengupta, R., Rangarajan, A. and Ranka, S.

Subcycle-based Neural Network Algorithms for Turning Movement Count Prediction.

DOI: 10.5220/0010501407350744

In Proceedings of the 7th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS 2021), pages 735-744

ISBN: 978-989-758-513-5

735

in on accurately and efﬁciently deriving TMCs using

loop detector data by utilizing machine learning tech-

niques.

So far, intersection systems have only given data

at coarse levels of granularity (for example, trafﬁc

movement counts by hour) limiting their use. The

recent advent of automated trafﬁc signal performance

measures (ATSPM) systems provides this information

and signal timing information at a decisecond level.

Availability of this information along with the avail-

ability of cheap GPU-based computing and deep neu-

ral network algorithms has opened up the possibility

of modeling trafﬁc behavior at the subcycle level and

is the focus of this paper.

The rest of the document is described as follows.

Section 2 presents the related work of different tech-

niques used for turning movement counts prediction.

Section 3 describes how we preprocess raw data and

generate synthetic datasets from real-world controller

log data using SUMO (Lopez et al., 2018). In sec-

tion 4, we present the machine learning models devel-

oped for predicting turning movement counts and sev-

eral experimental results. Conclusions are provided in

Section 5.

Our contributions can be summarized as follows:

• We show that using detector waveforms instead

of aggregated trafﬁc volumes leads to better accu-

racy for turning movement counts prediction. Our

neural network models perform well both under

unsaturated and oversaturated conditions.

• We have developed neural network models for

predicting turning movement counts for an inter-

section, given information from stop bar detectors

and advance detectors of the intersection and ad-

vance detectors of upstream intersections. Our ap-

proach uses a novel modular neural network archi-

tecture for predicting turning movement counts.

By using a common weight matrix for all four di-

rections in the ﬁrst hidden layer of the network,

the size of this network is much smaller than a typ-

ical feed-forward network, thereby reducing the

uncertainty in predictions.

• We have developed a system for generating syn-

thetic datasets with trafﬁc distributions close to

real world. We used detector controller logs from

300 intersections in Orlando to construct inﬂow

waveforms and fed them into SUMO to run simu-

lations.

2 RELATED WORK

In this section, we summarize the existing literature

on turning movement counts prediction. We broadly

categorize the existing literature and summarize each

category.

2.1 Flow-based Methods

Some studies, (Chen et al., 2012; P. T. Martin, 1992;

Wu and Thnay, 2001; Xu et al., 2013), have used

mathematical techniques like linear programming to

impute turning movement counts which balances in-

ﬂow, outﬂow of trafﬁc with some constraints. By

modelling the ﬂow with some constraints as linear

equations, we can solve for every unknown movement

if we have enough equations. To calculate westbound

through volume at intersection i, the following equa-

tion can be used.

W BT

= −NBL

− SBR

+W BR

t+δt

+W BL

t+δt

+W BT

t+δt

− δ

i j

(1)

where

WBT: westbound through

NBL: northbound left

SBR: southbound right

WBR: westbound right

WBL: westbound left

t: time interval t

i,j: intersections number

∆t time taken to travel between i,j

t+∆t

i j

: additional trips generated between inter-

sections i, j during time interval ∆t.

In the equation, ∆t is assumed to be zero if the dis-

tance between the two intersections is small because

the volume ﬂuctuations would be signiﬁcantly low.

Also, δ

i j

can be assumed to be zero if there are no

trips generated between these two intersections.

Linear programming-based approaches

(P. T. Martin, 1992) have been used to compute

turning movement counts using ﬂows. This approach

requires measurements like detector ﬂow, weight

given to each link, and constraints on ﬂow for each

link. A system of linear equations is modelled with

the above constraints and information. This system of

equations is solved to get turning movement counts.

Methods based on origin-destination (O-D) ma-

trices to compute turning movement counts (Wu and

Thnay, 2001) uses already observed O-D matrices to

predict future O-D matrices. They use a travel de-

mand model, the Furness Method, and then use the

equilibrium principle to assign future O-D matrices

to the network.

iMLTrans 2021 - Special Session on Intelligent Mobility, Logistics and Transport

736

Stopbar detector

Advance detector

Outﬂow detector

Figure 1: Diagram showing a typical intersection. Arrival/departure waveforms are observed at stop bar (S) and advance (A)

detectors.

The path ﬂow estimator (Chen et al., 2012) com-

putes complete link ﬂows along with turn movement

counts when some trafﬁc counts at selected intersec-

tions are known. The proposed algorithm iterates over

entropy and equilibrium terms to obtain the ﬁnal solu-

tion. The ﬁrst term, entropy, distributes trips to multi-

ple paths. The second term, equilibrium, makes those

trips cluster together on minimum cost paths. The al-

gorithm has an outer loop to iteratively generate paths

to the working path set as needed to replicate the ob-

served link counts, turning movement counts, and se-

lected prior O-D ﬂows. The claim is that the method

has advantages because of the single level convex

programming formulation, when compared to other

bilevel programming approaches.

2.2 Neural Network-based Approaches

These methods (Ghanim and Shaaban, 2019) use only

approach volumes to predict corresponding turning

movement counts. Figure 2 shows how a four-leg in-

tersection is represented as a node where the number

of vehicles leaving the node i is P

: P

∑

j=1

i j

. The

turning movement counts between node i and node j

are the number of trips from i to j. So, when a vehicle

Figure 2: Representation of signalized intersection pro-

posed in (Ghanim and Shaaban, 2019).

is approaching an intersection, they have four possible

turns: left, through, right, and U-turn. For the four ap-

proaches, there are 16 possible movements. For inter-

section j, the number of vehicles attracted to it is A

∑

i=1

i j

; this is the sum of all the trips with des-

tination j. So 16 different turning movement counts

should be modelled with the help of known values.

Each T

i j

is modelled as a function of P

and A

. (Lv

Subcycle-based Neural Network Algorithms for Turning Movement Count Prediction

737

et al., 2015) uses a stacked autoencoder model to pre-

dict trafﬁc ﬂow. An autoencoder is a neural network

that tries to reproduce its input. It acts as an iden-

tity function. It tends to learn features that form a

good representation of its input. Stacked autoencoder

models are created by stacking autoencoders in a hi-

erarchical way. The generic trafﬁc ﬂow features are

modelled with a stacked autoencoder model. They

also take into account spatial and temporal correla-

tions. They use a logistic regression layer as the top

layer for supervised trafﬁc prediction.

2.3 Other Approaches

Approaches such as the genetic algorithm-based

framework (Pengpeng Jiao, 2005) are also used to ob-

tain a dynamic relation between turning movement

proportions and trafﬁc counts at an intersection at

each time step. The proposed objective function min-

imizes the sum of absolute differences between ob-

served and predicted trafﬁc counts. The objective

function is made to converge using a revised pa-

rameter optimization model. Then, they ﬁnally pro-

pose a genetic algorithm to impute turning movement

counts.

The current state of research is lacking in several

aspects. The datasets used are often small and don’t

represent a wide variety of trafﬁc scenarios that may

occur. Current methods use data aggregated at res-

olutions in the order of minutes and above and thus

miss important details that may be found in more ﬁne-

grained data.

Our approach is different for two reasons:

1. Large number of samples collected using a realis-

tic simulation framework

2. Results show that using more ﬁne-grained data

results in better accuracy for turning movement

count prediction.

3 DATASETS AND

PREPROCESSING

In this section, we describe the data preprocessing and

dataset generation pipeline. We ﬁnd that the present

datasets on turning movement counts are inadequate

for our needs. Hence, we use the SUMO microscopic

simulator programmed with real-world trafﬁc ﬂows

and signal plans to generate our dataset. Speciﬁcally,

using SUMO helps us in the following ways:

• Datasets are often aggregated at 30-minute inter-

vals, which is too coarse for our needs. SUMO

allows us to generate data at 1-second resolution.

• Actual (ground truth ) turning movement counts

are not readily available nor is there a straightfor-

ward way of using controller log data for imputing

them. SUMO logs allow us to get ground-truth

turning movement counts.

• Another challenge is that we do not have a detec-

tor channel to acquire phase mappings for most of

the intersections. So, the information from con-

troller logs cannot be used in a structured manner

(we do not know whether a detector belongs to a

major or minor street or is advance or at a stop

bar). Because we place the detectors in SUMO,

we know their locations and functions precisely.

One important aspect of simulations is modeling

the demand to be as realistic as possible. Program-

ming random ﬂows or using coarse origin-destination

matrices for generating ﬂows may lead to unrealis-

tic trafﬁc distributions. The real-world dataset we use

to program our simulation consists of controller log

data from 329 signalized intersections in Seminole

County, Greater Orlando Metropolitan Area. For traf-

ﬁc ﬂows, we use loop detector waveform data from

advance detectors of these intersections and regener-

ate them in our simulations. We create a huge syn-

thetic dataset of 30 million cycles, with trafﬁc dis-

tributions similar to those observed in the real world.

(The process is described in 3.2.)

3.1 Recorded High Resolution Loop

Detector Data

High-resolution loop detector data and signal tim-

ing data captured by ATSPM (Day et al., 2014) is

valuable for characterizing the performance of in-

tersections using various Measures of Effectiveness

(MoEs). Induction loop detectors attached to the in-

tersection collect data at 10 Hz, indicating whether a

vehicle passed over or not. Signal behavior is also

captured. These data allow trafﬁc engineers to ana-

lyze the performance of trafﬁc intersections and im-

prove safety and efﬁciency while cutting costs and

congestion.

Table 1 shows a sample of high resolution data.

The data consist of the following attributes:

1. SignalID: Intersection identiﬁer

2. Timestamp: Time at which event was logged (de-

cisecond resolution)

3. EventCode: What event at the signal was captured

4. EventParam: What was the value of the event or

attribute at that timestamp.

These data also come with metadata which de-

scribe what different event codes and event param-

iMLTrans 2021 - Special Session on Intelligent Mobility, Logistics and Transport

738

Table 1: Raw event logs from signal controllers. Most mod-

ern controllers generate these data at a frequency of 10 Hz.

SignalID Timestamp EventCode EventParam

1490

2018-08-01

00:00:00.000100

82 3

1490

2018-08-01

00:00:00.000300

82 8

1490

2018-08-01

00:00:00.000300

0 2

1490

2018-08-01

00:00:00.000300

0 6

1490

2018-08-01

00:00:00.000300

46 1

1490

2018-08-01

00:00:00.000300

46 2

1490

2018-08-01

00:00:00.000300

46 3

eters indicate, for example, event code 81 indicates a

vehicle departure, event code 2 indicates start of green

phase, etc. The event parameter identiﬁes the partic-

ular detector channel or phase in which the event was

captured.

3.2 Data Generation

In order to model an intersection, we need a large

dataset to train our models. While several real-world

datasets exist (Wang et al., 2019), they are of limited

utility for our task because they do not model inter-

sections under diverse trafﬁc and signal timing condi-

tions.

Along with diverse but realistic trafﬁc ﬂows, we

also need to program realistic trafﬁc signal plans. In

the real world, it is highly unlikely that a trafﬁc au-

thority would implement undesirable signal timing

schemes for gathering data because that would have

adverse real-world consequences. On the other hand,

microscopic simulators offer us the ﬂexibility of im-

plementing undesirable signal timing plans. We study

the Orlando dataset and derive a signal timing plan as

shown in Table 2.

The signal timing plan for the intersection is an

actuated signal timing plan with minimum and maxi-

mum times, as shown in Table 2.

A yellow time of 5 seconds follows each phase.

Thus, this leads to a theoretical minimum signal cycle

length of 60 seconds and a maximum length of 180

seconds. This maximum is chosen with consideration

of acceptable pedestrian wait times.

We use a three-stage approach for generating sim-

Table 2: Actuated Signal Timing Plan Details.

Trafﬁc Min Green Max Green

Movement Time (seconds) Time (seconds)

Corridor Through/Right 20 70

Side-Through/Right 10 30

Corridor Left 10 30

Side Left 10 30

ulated data:

1. Generate a realistic intersection conﬁguration in

SUMO

2. Derive trafﬁc waveforms from real data

3. Run parallel simulations using waveforms from

step-2 and intersection conﬁguration in step-1.

3.3 Intersection Conﬁguration

Our simulation consists of a one-intersection sce-

nario with four approaches, based on standard

NEMA (National Electrical Manufacturing Associa-

tion) (US Department of Transportation, 2008) phas-

ing. It consists of four through or right movements

and four left-turn movements, one of each movement

for the four approaches. Most urban arterials have an

exclusive left-turn buffer at each approach to cater to

the left-turning trafﬁc. This prevents the left-turning

trafﬁc from blocking the through and right trafﬁc until

the buffer is ﬁlled.

Each approach is initially a single lane which

fans out into a through-lane and an exclusive left-turn

buffer. The left-turn buffers extend 60 meters and can

hold 6-7 vehicles. There are two stop bar detectors

per approach, one each for through/right and left turn

lanes. There is one advance detector 90 meters from

the intersection, just beyond the end of the left-turn

buffer. This minimal conﬁguration captures all the

eight turn movements possible. Multiple lanes for

each movement group can be handled by (a) aggre-

gating detector counts per movement group and (b)

training multiple models, one for each intersection ge-

ometry of interest. In this study, we only focus on the

most general and minimal conﬁguration.

In order to gather downstream data, we place gat-

ing trafﬁc signals that mimic a downstream intersec-

tion 800 meters from the main intersection. They

are simple one-phase signals without any side streets.

Adding fully-ﬂedged intersections with coherent real-

world ﬂows would have been computationally expen-

sive and will be included in a follow-up extension to

this work. There are four such gating signals, one

along each of the four outbound directions.

Subcycle-based Neural Network Algorithms for Turning Movement Count Prediction

739

3.4 Input Trafﬁc Generation

Archived controller logs are used to construct inﬂow

waveforms along all four directions. These wave-

forms are in turn used to inject vehicles into SUMO

from four directions with an adaptive signal timing

control. This way, we generate datasets whose inﬂow

and outﬂow distributions are close to what we observe

in the real world. For this, we used six months of con-

troller logs from 90 intersections in the City of Or-

lando.

Thus, the main ﬂow along corridor through and

right directions will be between two to eight times the

ﬂow along the other streets. These ratios are based

on the observed trafﬁc ﬂows in the recorded Orlando

dataset.

We use advance detector logs from the Orlando

dataset to generate vehicle ﬂows at a 1-second res-

olution. We randomly sample ﬂow patterns ob-

served at these two detectors for the straight and side

streets and ensure that they fall between the above-

mentioned volume ﬂow constraints. We program

these arrival patterns into SUMO. These patterns are

further shaped by the gating signals at the start of the

four incoming approaches. This ensures variable pla-

tooning of volumes based on real-world data.

3.5 Parallel Dataset Generation

The data generation process makes use of the multi-

processing environment. At any instant, several simu-

lations will be running in parallel: each thread runs a

simulation, processes the logs, and dumps the dataset

into the ﬁle system.

Each simulation generates logs which have infor-

mation of every time step of the simulation. These

logs are processed, and the following information is

stored:

• Waveforms at all the stop bar and advance detec-

tors for all the approaches

• Waveforms at all the advance detectors of nearby

intersections

• Signal timing information

• Turn movement counts for all the possible move-

ments.

Within a simulation, after an initial simulation

warm-up of 600 seconds, logs are extracted in win-

dows of 1,000 seconds. These usually contain 8-9

complete cycles on average. Thus, each data exem-

plar consists of a set of waveforms of different sig-

nals and detectors, queue lengths, and turn movement

counts for a window of 1,000 seconds (T = 200), ag-

gregated at 5-second resolution.

A large dataset of 5 million such exemplars is thus

generated, accounting for 30 million trafﬁc cycles of

simulation. The dataset is then split for training and

testing in the ratio of 70:30.

The creation of such a vast dataset involved con-

siderable engineering effort. The entire pipeline was

implemented in the Python programming language. A

multiprocessing library was used to run up to 60 par-

allel instances of SUMO and preprocess output XML

logs in batches. Numpy (Harris et al., 2020) and

Dask (Rocklin, 2015) were then used to create vectors

for training and testing. These vectors were stored

in HDF5 format using the H5PY (Collette, 2013) li-

brary.

Implementation, training, and evaluation of the

deep learning models was done using the Py-

Torch (Paszke et al., 2019) library. University of

Florida’s HiPerGator supercomputing resources were

used to train and test multiple models in parallel.

We use this dataset for modelling a universal ap-

proximator for predicting turn movement counts. The

inputs to the model are:

• Detector actuations for all advance detectors (four

detectors)

• Detector actuations at stop bar detectors for

through-right and left buffers (eight detectors)

• Detector actuations at early detectors of nearby in-

tersections (outﬂow detectors) (four detectors).

The information from these 16 detectors is used to

predict turn movement counts. The detector actua-

tions are aggregated to some level to construct detec-

tor waveforms. These detector waveforms are in turn

used as input to the neural network models.

4 MACHINE LEARNING

MODELS FOR TMC

PREDICTION

The input to the models is detector waveforms (ag-

gregated to some time interval) at stop bar and

advance detectors. The input layer has 16 ×

(numbero f timesteps) neurons, each neuron takes in-

put from one detector and one time point.

The waveforms are represented as 1-D vectors,

each with T components. Here T refers to the length

of time a particular sensor’s data is being considered,

with each component being aggregated at a 5-second

level.

We performed different experiments to analyze

how the following parameters affect the TMC predic-

tion.

iMLTrans 2021 - Special Session on Intelligent Mobility, Logistics and Transport

740

Figure 3: Simulations block diagram.

Figure 4: Turning movement count distributions of the dataset generated by simulation.

• Prediction Window: This deﬁnes the time win-

dow for which we predict the TMCs. Based on

our results, we observed that as the size of the pre-

diction window increases, the accuracy for pre-

dicting turning movement counts increases.

• Waveform Aggregation Level: This denotes the

time resolution for each step of the waveform con-

structed from detector actuations. We vary this to

ﬁnd the optimal value of aggregation level for the

waveform (5 sec, 15 sec, 25 sec, etc.).

We now describe a modular neural network archi-

tecture where input is transformed to the ﬁrst hidden

layer by sharing a common weight matrix among the

detector waveforms of all the directions. The basic

reasoning is that the detector waveforms (stop bar and

advance) of the four directions would have similar

properties. So, for the feed-forward, network we ana-

Table 3: Table showing MSE on test set for prediction win-

dow of 375 seconds for different waveform aggregation lev-

els.

Predicton

Window

Waveform Aggrega-

tion level

MSE on

Test Set

375 sec 5 sec 0.0002

375 sec 15 sec 0.0005

375 sec 25 sec 0.0002

375 sec 75 sec 0.0004

375 sec 375 sec 0.001

lyze the weights connecting the inputs to ﬁrst hidden

layer. Figure 10 shows the heat map of weights con-

necting the input layer to the ﬁrst hidden layer of the

feed-forward network. In the plot the absolute value

of weights is taken, multiplied by 100 to highlight the

trends. The red marker in the plot separates the four

Subcycle-based Neural Network Algorithms for Turning Movement Count Prediction

741

Figure 5: Actual vs. predicted turning movement counts for six different movements using a waveform model. The y-axis

represents particular movements; the x-axis represents random exemplars of the test set.

Table 4: Table showing MSE on test set for different pre-

diction windows.

Prediction

Window

Waveform Aggrega-

tion level

MSE on

Test Set

375 sec 5 sec 0.0002

375 sec 15 sec 0.0005

750 sec 5 sec 0.0005

750 sec 15 sec 0.0006

750 sec 25 sec 0.0005

1000 sec 5 sec 0.0008

1000 sec 20 sec 0.0008

Figure 6: Plot showing MSE for eastbound through trafﬁc

for different bins, comparing different waveform aggrega-

tion levels.

directions, and the detector waveforms on the y-axis

are ordered in correspondence to detectors in this or-

der: east, west, north, and south. We can see from the

plot that the weights connected to detectors on differ-

ent directions look similar.

So to decrease the number of learnable parameters

and get better generalization, a common weight ma-

Figure 7: Plot showing percentage error for comparing dif-

ferent bins. This plot suggests that the model is perform-

ing well irrespective of the saturation level: (a) southbound

through; (b) eastbound through.

trix is used for transforming detector waveforms on

different directions to the ﬁrst hidden layer. Figure 9

shows the comparison of percentage error for differ-

ent bins comparing a fully connected network with

the modular architecture. We can see that the mod-

ular architecture performs equally well with a lower

number of parameters.

5 CONCLUSIONS

We have developed machine learning models for pre-

dicting turn movement counts for a single intersec-

tion, given information from stop bar and advance de-

tectors. All of these data are available from ground

sensors. Hence, this model can be easily adapted to

practical datasets.

iMLTrans 2021 - Special Session on Intelligent Mobility, Logistics and Transport

742

(a)

(b)

Figure 8: Plot showing percentage error for comparing dif-

ferent prediction windows. This plot suggests that accuracy

increases when predicting for a larger window size: (a) east-

bound through; (b) southbound through.

(a)

(b)

Figure 9: Plot showing comparison between fully con-

nected and weight-sharing model. The weight-sharing

model is comparable or better.

• We developed neural network models for predict-

ing turning movement counts for an intersection,

given information from stop bar and advance de-

tectors of the intersection and advance detectors

of upstream intersections.

Figure 10: Plot showing heat map of weights connecting in-

put layer and ﬁrst hidden layer.The horizontal black bar sep-

arates the four directions. The ordering of detector wave-

forms along the y-axis is as follows: east, west, north, and

south.

• We developed a system for generating synthetic

datasets with trafﬁc distributions close to real-

world. We used detector controller logs from

300 intersections in Orlando to reconstruct inﬂow

waveforms and fed them into SUMO for running

simulations.

• We prove that using detector waveforms instead

of aggregated trafﬁc volumes leads to better accu-

racy for predicting turning movement counts.

• The neural network models developed perform

well both under unsaturated and oversaturated

conditions.

• We developed a modular neural network archi-

tecture for predicting turning movement counts.

Instead of using different weight matrices across

four directions (as in general feed-forward net-

works), we use a common weight matrix for all

Subcycle-based Neural Network Algorithms for Turning Movement Count Prediction

743

the four directions in the ﬁrst hidden layer of

the network. This modular architecture performs

equally well and has fewer learnable parameters

in comparison with a feed-forward network.

ACKNOWLEDGEMENTS

The work was supported in part by NSF CNS

1922782 and by the Florida Department of transporta-

tion (FDOT) District 5. The opinions, ﬁndings and

conclusions expressed in this publication are those of

the authors and not necessarily those of FDOT D5.

REFERENCES

Chen, A., Chootinan, P., Ryu, S., Lee, M., and Recker, W.

(2012). An intersection turning movement estimation

procedure based on path ﬂow estimator. Journal of

Advanced Transportation, 46:161 – 176.

Collette, A. (2013). Python and HDF5. O’Reilly.

Day, C. M., Bullock, D. M., Li, H., Remias, S. M., Hainen,

A. M., Freije, R. S., Stevens, A. L., Sturdevant, J. R.,

and Brennan, T. M. (2014). Performance measures for

trafﬁc signal systems: An outcome-oriented approach.

Technical report.

Ghanim, M. S. and Shaaban, K. (2019). Estimating turn-

ing movements at signalized intersections using artiﬁ-

cial neural networks. IEEE Transactions on Intelligent

Transportation Systems, 20(5):1828–1836.

Harris, C. R., Millman, K. J., van der Walt, S. J., Gommers,

R., Virtanen, P., Cournapeau, D., Wieser, E., Taylor,

J., Berg, S., Smith, N. J., Kern, R., Picus, M., Hoyer,

S., van Kerkwijk, M. H., Brett, M., Haldane, A., del

R’ıo, J. F., Wiebe, M., Peterson, P., G’erard-Marchant,

P., Sheppard, K., Reddy, T., Weckesser, W., Abbasi,

H., Gohlke, C., and Oliphant, T. E. (2020). Array pro-

gramming with NumPy. Nature, 585(7825):357–362.

Lopez, P. A., Behrisch, M., Bieker-Walz, L., Erdmann, J.,

otter

od, Y.-P., Hilbrich, R., L

ucken, L., Rummel, J.,

Wagner, P., and Wießner, E. (2018). Microscopic traf-

ﬁc simulation using sumo. In The 21st IEEE Interna-

tional Conference on Intelligent Transportation Sys-

tems. IEEE.

Lv, Y., Duan, Y., Kang, W., Li, Z., and Wang, F. (2015).

Trafﬁc ﬂow prediction with big data: A deep learn-

ing approach. IEEE Transactions on Intelligent Trans-

portation Systems, 16(2):865–873.

P. T. Martin, M. C. B. (1992). Network programming to

derive turning movements from link ﬂows. In Transp.

Res. Rec., no. 1365, pages 147–154. Transportation

Research Board.

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J.,

Chanan, G., Killeen, T., Lin, Z., Gimelshein, N.,

Antiga, L., Desmaison, A., Kopf, A., Yang, E., De-

Vito, Z., Raison, M., Tejani, A., Chilamkurthy, S.,

Steiner, B., Fang, L., Bai, J., and Chintala, S. (2019).

PyTorch: An Imperative Style, High-Performance

Deep Learning Library.

Pengpeng Jiao, Huapu Lu, L. Y. (2005). Real-time esti-

mation of turning movement proportions based on ge-

netic algorithm. In Proceedings. 2005 IEEE Intelli-

gent Transportation Systems, 2005., pages 96–101.

Rocklin, M. (2015). Dask: Parallel computation with

blocked algorithms and task scheduling. In Huff,

K. and Bergstra, J., editors, Proceedings of the 14th

Python in Science Conference, pages 130 – 136.

US Department of Transportation, F. H. A. (2008). Traf-

ﬁc signal timing manual. https://ops.fhwa.dot.gov/

publications/fhwahop08024/chapter4.htm. (Accessed

on 03/18/2021).

Wang, Y., Zhang, D., Liu, Y., Dai, B., and Lee, L. H. (2019).

Enhancing transportation systems via deep learning:

A survey. Transportation research part C: emerging

technologies, 99:144–163.

Wu, J. and Thnay, C. (2001). An o-d based method for

estimating link and turning volume based on counts.

In Transp. Res. Rec., no. 1365, pages 865–873. ITE

Dist 6.

Xu, K., Yi, P., Shao, C., and Mao, J. (2013). Development

and testing of an automatic turning movement identi-

ﬁcation system at signalized intersections. Journal of

Transportation Technologies, 03:241–246.

iMLTrans 2021 - Special Session on Intelligent Mobility, Logistics and Transport

744