Road Traffic Modeling and Management: Using Statistical Monitoring and Deep Learning

By Fouzi Harrou, Abdelhafid Zeroual, Mohamad Mazen Hittawe and Ying Sun

Road Traffic Modeling and Management: Using Statistical Monitoring and Deep Learning provides a framework for understanding and enhancing road traffic monitoring and management. The book examines commonly used traffic analysis methodologies as well the emerging methods that use deep learning methods. Other sections discuss how to understand statistical models and machine learning algorithms and how to apply them to traffic modeling, estimation, forecasting and traffic congestion monitoring. Providing both a theoretical framework along with practical technical solutions, this book is ideal for researchers and practitioners who want to improve the performance of intelligent transportation systems.

• Provides integrated, up-to-date and complete coverage of the key components for intelligent transportation systems: traffic modeling, forecasting, estimation and monitoring
• Uses methods based on video and time series data for traffic modeling and forecasting
• Includes case studies, key processes guidance and comparisons of different methodologies

• Language: English
• Publisher: Elsevier Science
• Release date: Oct 5, 2021
• ISBN: 9780128234334

Fouzi Harrou

Fouzi Harrou received the M.Sc. degree in telecommunications and networking from the University of Paris VI, France, and the Ph.D. degree in systems optimization and security from the University of Technology of Troyes (UTT), France. He was an Assistant Professor with UTT for one year and with the Institute of Automotive and Transport Engineering, Nevers, France, for one year. He was also a Postdoctoral Research Associate with the Systems Modeling and Dependability Laboratory, UTT, for one year. He was a Research Scientist with the Chemical Engineering Department, Texas A&M University at Qatar, Doha, Qatar, for three years. He is actually a Research Scientist with the Division of Computer, Electrical and Mathematical Sciences and Engineering, King Abdullah University of Science and Technology. He is the author of more than 150 refereed journals and conference publications and book chapters. He is co-author of the book "Statistical Process Monitoring Using Advanced Data-Driven and Deep Learning Approaches: Theory and Practical Applications" (Elsevier, 2020). Dr. Harrou’s research interests are in the area of statistical anomaly detection and process monitoring with a particular emphasis on data-driven, machine learning/deep learning methods. The algorithms developed in Dr. Harrou’s research are utilized in many applications to improve the operation of various environmental, chemical, and electrical systems.

Book preview

Chapter 1: Introduction

Abstract

The management of traffic congestion in public road networks is becoming a key factor for economic growth. Indeed, road traffic congestion hampers economic growth: it has a profound impact on travel time, causing delays that result in the significant financial loss of billions of dollars spent on extra hours of travel and fuel consumption. In this chapter, we present different sensors technologies commonly employed to collect traffic flow data, including intrusive and non-intrusive sensors. Then we describe the basic traffic features commonly used to manage and control traffic congestion. After that, we briefly overview the commonly used traffic models in the literature for modeling and forecasting traffic flow, including data-driven and model-driven models. Finally, we review the objective and structure of the book.

Keywords

Traffic flow forecasting; Road traffic sensor; Traffic features; Traffic flow modeling; Intelligent transportation systems

1.1 Introduction

Recently, vehicular traffic has been significantly increased, particularly with urbanization and expansion of cities. The highly populated cities are often exposed to unbearable levels of traffic congestion on urban roads. Essentially, traffic congestions occur when the demand surpasses the available roadway infrastructures [1]. Traffic congestion can be characterized by high vehicle density, which leads to an excess of travel time [2]. Two kinds of traffic congestions can be distinguished depending on the reasons causing congestion, recurring and non-recurring congestions [3]. Crucially, recurring congestion is mainly due to excess demand and the absence of sufficient infrastructures [4]. However, non-recurring congestion is principally generated because of incidents, unsuited weather conditions, and work areas [5].

In recent years, there has been a significant increase in vehicles, causing traffic congestion, which has become a considerable problem from safety, economic, and environmental perspectives. Crucially, traffic congestion leads to increased travel times, inflicting costs on the economy, and also causes various influences on urban zones and their residents [6]. Thus the urban transportation systems have been negatively impacted by the travel delay and costs induced via traffic congestion. Specifically, the available infrastructures are incapable of accommodating the growing number of vehicles, which generated increased traffic congestion levels. As pointed in [7], in the US the trend of traffic congestion from 1982 to 2019 is persistently increased by every measure. Congestion traffic levels dramatically declined in 2020 due to the COVID-19 pandemic and lockdowns, but the trends in the last few months of 2020 point to a return of congestion problems in 2021 [7]. For instance, in the US, traffic congestion in 2014 caused 6.9 billion extra hours, leading to purchasing 3.1 billion additional gallons of fuel, costing $160 billion [8]. In addition, it is reported in [9] that $480 billion has been spent by drivers in the most congested 25 cities of the US because of time delay, consumed fuel, and emitted carbon under congestion traffic. The US total cost of lost productivity because of traffic congestions is estimated to be $87 billion in 2018 [10].

A continuous increase in road traffic demand causes several problems, including traffic congestion, air pollution, and accidents that can cause severe injuries and even deaths. Despite increased congestions in recent years, promising improvements have been developed in Intelligent Transportation Systems (ITS) to handle traffic congestions efficiently. Importantly, the main objective behind ITSs is to achieve safe, effective, and environment-friendly transportation systems by exploiting advanced technologies and creating cooperation linking users, vehicles, and road infrastructures. Specifically, ITS aim to mitigate traffic congestion, minimize travel time, and create safe transport networks. To this end, ITS come up with different mart traffic management systems to manage traffic flow, including smart traffic lights [11,12], ramp metering [13], variable speed limits [14], and specialized sensing technologies.

Numerous ITS solutions have been developed to mitigate road safety issues by creating more intelligent interactions between roads, vehicles, and users. For instance, the cooperative collision warning system is introduced for alerting drivers on the presence of potential collision risks by using information obtained from surrounding vehicles and road infrastructure [15]. There are several collision warning systems to prevent vehicles and improve road safety [16–20]. Several ITS technologies were designed in various countries. For example, one remarkable ITS technology was during the 2010 Asian Games in China for parallel traffic control and management [21]. Of course, the central role of ITS is to manage and reduce congestion using online measurements about traffic patterns, including density, speed, travel time, and the position of vehicles.

The accurate short-term forecasting of traffic flow is essential in developing intelligent transportation systems (ITS), especially to efficiently manage traffic flow and mitigate traffic congestion [22]. Given reliable traffic flow forecasting, travelers and different government sectors use such information to optimally deal with traffic congestion by using more appropriate travel routes, evaluate roadway safety, reduce pollution, and estimate energy consumption [23].

1.1.1 Types road traffic sensors

Management of road networks is based on tracking the evolution of the traffic features defined above. The manager's objective is to collect as much information as possible about the network operating state to efficiently manage traffic flow. To this end, different sensor technologies are available to gather raw traffic data [24]. Generally, we distinguish two types of sensors, in-vehicle sensors and in-road sensors (Fig. 1.1).

Figure 1.1 Types of sensors technology in ITS.

In recent years, several sensors have been embedded in vehicles to improve drivers' comfort and safety, including LIDAR (Light Detection And Ranging), RAdio Detection And Ranging (RADAR), laser sensors, gyroscope and accelerometer sensors, and Global Positioning Systems (GPS). For instance, the use of RADAR and laser sensors enable avoiding collisions by scanning the road for frontal side and rear collisions and allowing safety applications. These in-vehicle sensors provide relevant information to reduce traffic congestion, facilitate parking, and decrease road accidents [25].

The second category of sensors frequently used by the ITS consists of in-road sensors. Even with significant improvement in the automotive industry for improving safety and comfort in vehicles, traffic measurements collected from sensors placed on the roadside are becoming challenging for ITSs. These in-road sensors offer pertinent information to users, including smart parking and traffic congestion levels on the road. In addition, these sensors gather weather data, which is very useful to improve transportation networks and traffic management. Globally, sensors within the in-road class can be divided into two sub-categories based on their location, intrusive sensors and non-intrusive sensors [26]. Intrusive sensors are installed on road surfaces. Several intrusive sensors are used in practice, including pneumatic road tubes, inductive loop detectors (ILD), magnetic sensors, and piezoelectric sensors [24]. They are characterized by high accuracy and rare missing data because they are broadly installed and have high precision in sensing vehicles. Moreover, the key features of traffic are all measured, possibly by iterating two short-distance sensors to estimate the individual speed. However, their major drawback consists of the high cost of their installation and maintenance.

Non-intrusive sensors are placed on infrastructure outside the roadway. They can sense vehicles and other traffic features, such as speed and lane occupancy. However, they are sensitive to environmental conditions. Non-intrusive sensors include video cameras, Radar sensors, infrared, ultrasonic, and acoustic array sensors. Essentially, non-intrusive sensors are employed to give information on a chosen area, like detecting queues at a traffic light and providing weather conditions of the monitored road.

Traffic sensors can be also be classified into two classes, perception and non-perception sensors. Essentially, perception sensors consist of vision-based sensors, such as cameras, LiDARs, and RADARs. These sensors gather images and/or videos of traffic scenes. We can find these sensors installed onboard vehicles or on the roadside. On the other hand, non-perception sensors collect traffic features like flow, density, vehicle location, and internal measurements. Non-perception sensors comprise in-vehicle sensors (e.g., GPS and Controller Area Network (CAN) bus) and in-road sensors (e.g., inductive loop counter).

1.1.2 Key traffic features

This section presents preliminary information about traffic; for more details, see, for instance, [27,28]. Traffic flow dynamics of vehicles on the road network are of central interest to the infrastructure managers and decision-makers of transportation. Four main variables are traditionally used to characterize traffic dynamics in the network at a macroscopic level: traffic flow q, speed v, density ρ, and travel time. Transportation engineers can investigate these features from microscopic and macroscopic levels, but the focus here is on the macroscopic level. The planners exploit traffic flow measurements to monitor traffic flow, predict traffic trends, and plan and design facilities. Furthermore, traffic engineers employ these measurements for analyzing traffic operations and accidents and study improvements like priority lanes.

•  Traffic flow, also called the flow rate or volume, represents the number of vehicles between two points on the network during a given period T. In other words, it characterizes the intensity of the flow crossing a portion of the network and is expressed in vehicles per hour (veh/h).

•  Traffic speed, the second key traffic flow feature, is described as the distance traversed per unit time. It characterizes the velocity of the moving flow on the network in kilometers per hour (km/h).

•  Traffic density. We define the traffic density ρ as the number of vehicles on a given length of road (i.e., per unit space). This third fundamental characteristic is not directly measurable in practice; it can be estimated based on the measurements of the occupancy rate at a point in the traffic network. Of course, the traffic density is the number of vehicles that occupy a portion of the roadway. The traffic density plays an essential role in freeway control and monitoring systems.

•  Travel time (TT) represents the time needed by a vehicle in the flow to join two distinct network points. In other words, for any vehicle, the individual TT is defined by the difference in passage times between points A and B of the network.

Fig. 1.2 illustrates the space-time diagram, which is usually used to represent the dynamics of a portion of the road (length L) over a given period T from the trajectories of the vehicles.

Figure 1.2 Basic representations of macroscopic traffic variables.

In [29] a generalized definition of traffic states is provided, where a traffic state in a time-space region A is mathematically expressed as follows [29]:

(1.1)

(1.2)

(1.3)

where denotes the total distance traversed by all the vehicles in the area A (veh/km), refers to the total time spent by all the vehicles in area A (veh/h), and is the time-space area of the area A (km/h).

The traffic flow variables are linked by a phenomenological relationship called the fundamental diagram (FD) introduced by Greenshields [30]. It illustrates the empirical relationships between traffic state variables. Relevant information can be learned from an FD about traffic features, including free-flow speed and capacity. Different forms for FDs have been adopted to reflect certain relevant links associated with traffic (Fig. 1.3). A triangular FD is one of the most common forms due to its simplicity and some theoretically suitable characteristics [31,32].

Figure 1.3 Fundamental diagrams linking the macroscopic features of traffic according to the Greenshield model.

1.1.3 Traffic flow modeling

The need for traffic flow modeling and monitoring methods that can accurately model and detect traffic congestion has gained the attention of researchers and engineers. Over the past few decades, numerous traffic flow models have been introduced [33–37]. Globally, traffic flow modeling and monitoring techniques can be classified into two main categories, data-driven and model-driven techniques [37]. Traffic flow monitoring using model-based techniques can be grouped into two main categories, microscopic and macroscopic [33] (Fig. 1.4). Importantly, microscopic methods aim to model the behavior of the vehicle–conductor pair and its correlation with nearby vehicles [38–43]. However, approaches in this category require many parameters to reasonably model the traffic dynamics, which results in a high computational cost [33]. In contrast, macroscopic modeling methods rely on a hydrodynamic theory to model traffic dynamics as a continuous flow [44–47]. The major characteristic of macroscopic methods consists of using fewer parameters compared to microscopic modeling, which results in more parsimonious models with lower computational costs [34]. Crucially, the efficiency of these model-driven approaches relies on the accuracy of the models employed. The main drawback of these models is that they are more appropriate to small-scale traffic networks.

Figure 1.4 Traffic flow models.

On the other hand, data-driven methods rely on the availability of historical traffic flow data. These data are first used to build an empirical model, which is then used to detect traffic congestion in future data. Methods in this category comprise times-series models, machine learning models, and deep learning models. Traditional time-series models are among the most widely adopted in the literature for traffic flow modeling and forecasting. These methods include autoregressive integrated moving average (ARIMA) and its variants, such as Kohonen-ARIMA [48] and seasonal-ARIMA [49], and Holt–Winters models [50,51]. Parametric models, such as ARIMA and its variants, can achieve good performance when traffic flow time-series data show regular variations. However, the forecast error is apparent when the traffic flow shows irregular variations [52]. To alleviate this limitation, researchers paid much attention to non-parametric models. Over the last decade, shallow machine learning techniques, including neural network [53] and support vector machine [54], which are non-parametric and flexible models, are widely employed to improve traffic flow forecasting. They are characterized by convenient characteristics and their ability to model complex data without an analytical model formulation. For instance, in [55] a hybrid method combining a radial basis function neural network with the aid of a fuzzy system is proposed to forecast road traffic speed. A short-term traffic flow forecasting framework is proposed in [56] using least squares support vector regression; the authors study the impact of time lag on the quality of the forecasting. In [57] a hybrid model is established for traffic flow forecasting by combining complete ensemble empirical mode decomposition with adaptive noise and extreme gradient boosting. In [55] a hybrid model is introduced to forecast road speed by amalgamating the desirable features of the fuzzy logic system with the radial basis function (RBF) neural network. This hybrid RBF model has been trained using road network data and weather data to forecast the road speed and congestion status.

Deep learning has recently arisen as a promising research line in modeling and forecasting time series data, both in academia and industry [52,58–60]. Various deep techniques have been applied in the literature to address traffic flow and monitoring forecasting [61]. In [62] a stacked auto-encoder architecture is used to build a traffic flow prediction; the deep learning model training is based on a greedy layerwise approach. An ensemble learning approach is considered in [63] for traffic flow prediction based on a set of statistics and machine learning models. In [64] a data-driven approach for short-term passenger flow forecasting framework composed of several stages is proposed, where the authors employ a multi-variate approach for forecasting. In [65] a short-term forecast of traffic flow using the gated recurrent units is utilized with an analysis of the space-time.

1.2 Objectives and structure of the book

The book provides two main outputs, which include a theoretical framework and technical solutions. It aims to comprehensively provide an overview of traffic modeling, estimation, forecasting, and monitoring approaches (multivariate statistical techniques and deep learning approaches).

In Chapter 2, we present an overview of traffic modeling approaches, including microscopic, mesoscopic, and macroscopic modeling. More attention is given to macroscopic modeling to meet the aim of the book, traffic flow management. Both continuous and discrete versions of first-order macroscopic models are presented. Furthermore, the piecewise switched macroscopic traffic modeling approach, an enhanced version of the discrete cell transmission model, is studied. This model is a macroscopic traffic model that switches among a set of linear subsystems, with each subsystem representing traffic dynamically. Tests on traffic measurements, such as the four-lane State Route 60 and the three lanes Interstate 210 (I-210) in California freeways, are provided to show the prediction quality of the investigated methods.

In Chapter 3, we give an introduction to traffic data estimation. Then we present an observability study of the PWSL model, which is a necessary step to verify that the system can construct the traffic density state using reduced sensors information. Next, we deal with the traffic density estimation, in which we give an overview of the hybrid observer because of the hybrid nature of the PWSL model. After that, we present the PWSL-based hybrid observer beginning with its design and passing to its continuous and discrete estimating parts. Lastly, we present an alternative estimation approach based on the free PWSL Kalman filter, which is designed using only the normal traffic mode without any congestion. We apply it on data from road sections of the SR60 and I210 from the California highways to validate these approaches. Finally, we discuss the proposed approaches and obtained