2021

Navid Amoei khorshidi, Iran University of Science and Technology
Shahriar Afandizadeh Zargari, Iran University of Science and Technology
Navid Kalantari, AECOM

Travel time reliability is known as the temporal variability of travel time and influences many aspects of the traveler’s decisions, especially in roadway transport. This study aims to take a step in this direction by providing a tool to analyze the effect of transportation network improvements on travel time reliability. The primary issue that we address in this study is to find out how the reduction in the number of lanes, crashes, and traffic volume influence travel time reliability on interstate highways in Virginia. This study uses the data extracted from the INRIX for the interstate highways in the state of Virginia. We have examined different supervised data mining and Machine Learning (ML) algorithms to find the most promising tool and incorporated the results into an analysis toolbox. The results have shown to be satisfactory for modeling moderate travel time variations.

Corey Harper (cdharper@andrew.cmu.edu), Carnegie Mellon University
Sean Qian, Carnegie Mellon University
Constantine Samaras, Carnegie Mellon University

Heavy rainfall events are becoming more common in many areas with climate change, and these events affect travel speed and road safety. It is critical to understand how rainfall events affect congestion in the transportation network to help improve decision-making for infrastructure planning and real-time operations. This study incorporates high-resolution rainfall data into a travel speed prediction model to assess if localized rainfall data could inform travel speed prediction during light and heavy rainfall events, and how this compares to the classical method of utilizing a single city-wide rain gauge data point. Here we propose a data-driven approach that directly connects raw high-resolution rainfall data to travel speed, which avoids any complex hydrologic modeling. The travel speed prediction model holistically selects the most related features from a high-dimensional feature space by correlation analysis LASSO to overcome overfitting issues. The proposed method is applied to two urban arterials for case studies, located in Pittsburgh, PA. We find that in most cases, those rainfall features located relatively far away (i.e., greater than 5mi) from the segment of study have higher significance in predicting travel speed 30 minutes in advance. The results indicate that in our case study, it is beneficial to incorporate high-resolution rainfall data into a travel speed prediction model, but there may be certain days where high-resolution models do not outperform models that have a single city-wide rain gauge. This has implications for other cities that are interested in improving travel speed prediction modeling.

Satya Patra, Indian Institute of Technology, Madras
Bharathiraja Muthurajan
Bharathiraja Muthurajan
Lelitha Vanajakshi (lelitha@iitm.ac.in), Indian Institute of Technology, Madras

Travel time information is pivotal to several Intelligent Transportation Systems (ITS) applications. However, it is a dynamic and spatial variable that is difficult to measure using traditional sensors. It can be directly measured using aerial photography, tracking sensors or by vehicle re-identification. Aerial recording are difficult to manage, re-identification systems may not perform consistently under varying traffic conditions, and tracking systems needs user participation that restricts sample size. These shortcomings lead to investigating new technologies for travel time data collection. The proliferation of smartphones allowed researchers to explore the use of Wi-Fi Media access control Scanners (WMS) as a source of complementary travel time information. The current work is based on this concept and develops a low-cost WMS. It is then evaluated against a commercial counterpart for travel time data collection. Results show comparable or better performance with around 80% reduction in price. From this collected travel time information, predicting future values, which is of interest to many Intelligent Transportation Systems applications, is another challenging problem. Travel times are a result of interaction among many different constituent factors. The stochasticity of this traffic variable makes the application of any deterministic predictive model very demanding. Hence, constructing Prediction Intervals (PIs) than predicting a mean value is used in this study. The forecast results showed better performance when compared against Google's PI in all the studied routes.

Erik Jenelius (jenelius@kth.se), KTH Royal Institute of Technology

This paper investigates to what degree the vehicles traversing a route during the morning or evening peak over multiple days are recurrent travellers. The paper proposes a model of route speed distributions that separates the variability into a traveller-to-traveller component, consistent across days and time intervals, and a within-traveller component. The within-traveller component is further split into day-to-day, interval-to-interval and residual variability. Using data from Bluetooth and Wifi sensors on 26 routes in Stockholm, Sweden over a three-month period, we find that the traveller recurrence is higher towards the city in the morning peak and out from the city in the afternoon. Model estimation results show that the relative individual (within-traveller) variability is significantly higher in the commute direction (towards the city in the morning and out from the city in the afternoon) and on routes with high congestion levels. The results indicate that a distinction must be made between the within-traveller variability, which corresponds to travel time uncertainty, and the total variability which is typically captured by empirical measurements and used in travel time reliability assessments. Without this distinction, the costs associated with travel time variability may be overestimated.

Nima Hoseinzadeh (nhoseinz@vols.utk.edu), University of Tennessee, Knoxville
Yangsong Gu, University of Tennessee, Knoxville
Lee Han, University of Tennessee
Candace Brakewood, University of Tennessee, Knoxville
Brad Freeze, Tennessee Department of Transportation

In traffic operations, the aim of transportation agencies and researchers is typically to reduce congestion and improve safety. To attain these goals, agencies need continuous and accurate information about the traffic situation. Level-of-Service (LOS) is a beneficial index of traffic operations used to monitor freeways. The Highway Capacity Manual (HCM) provides analytical methods to assess LOS based on traffic density and highway characteristics. Generally, obtaining reliable density data on every road in large networks using traditional fixed location sensors and cameras is expensive and otherwise unrealistic. Traditional intelligent transportation systems (ITS) facilities are typically limited to major urban areas in different states. Crowdsourced data are an emerging, low-cost solution to potentially improve safety and operations. This study incorporates crowdsourced data provided by Waze to propose an algorithm for LOS assessment on an hourly basis. The proposed algorithm exploits various features from crowdsourced Waze user alerts and speed/travel time variation to perform LOS classification using machine learning models. Three categories of model inputs are introduced: basic statistical measures of speed, travel time reliability measures, and the number of hourly Waze alerts. Data collected from fixed location sensors were used to calculate ground truth LOS. The results reveal that using Waze crowdsourced alerts can improve LOS estimation accuracy by about 10% (accuracy=0.93, Kappa=0.83). The results of this research provide transportation agencies a LOS method based on crowdsourced data on different freeways segment (urban or rural) regardless of the availability of traditional fixed location sensors.

Xiaojian Hu, Southeast University
Tong Liu, Southeast University
Ye Li, Central South University
Zhiwei Cui, Head of the smart city institute
Tong Liu, Southeast University

ABSTRACT With the advancement of information technology and the popularization of navigation equipment, many ideas for research on traffic conditions have been provided. This paper obtains the average travel speed information of each road section in the road network through the Amap platform, and proposes a k-nearest neighbor model method suitable for this data type. According to the characteristics of the crawled data, the space-time two-dimensional state vector is used, and the Euclidean distance is used to measure the distance between the data. In order to improve the accuracy of the model, considering the difference of time and space dimension data, the two dimensions are weighted with exponential weighting method, correlation coefficient weighting, and Gaussian weighting, and the optimal weighting method is selected. The predicted value of the average travel speed of the road section is obtained through the prediction model. According to the threshold relationship between the speed and the traffic state, the traffic state level of the road section is judged to achieve the purpose of predicting the road traffic state. At the same time, the article also constructed a multi-step prediction k-nearest neighbor model based on the optimal weighting method, which achieved a higher accuracy rate of prediction within 30 minutes. Keywords: k-nearest neighbor, Spatiotemporal, Gaussian weighting, Euclidean distance

Homa Taghipour (htaghi2@uic.edu), University of Illinois, Chicago
Amir Bahador Parsa, University of Illinois, Chicago
Abolfazl Mohammadian, University of Illinois, Chicago

Accurate travel time prediction could improve highway traffic management and travelers’ decision making. In this study, three data driven techniques are employed to predict the travel time of highway links. For this purpose, Neural Network, K-Nearest Neighbors and Random Forest models along with multiple data sources including loop detectors, probe vehicles, weather condition, geometry, accidents, road works, special events, and sun glare are used. Then, using a dynamic approach and result of travel time prediction models, travel time of a highway corridor is calculated. Results confirm the outperformance of Random Forest model among three tested models. Furthermore, using the dynamic approach accuracy of travel time prediction for highway corridor improves by 4 percent, compared to snapshot travel time prediction approach.

Pouyan Hosseini, ITERIS, Inc.
Maan Qraitem, Boston University
Tiffany Symes, ITERIS, Inc.
Shayan Khoshmagham, ITERIS, Inc.

Short-term traffic speed predictions are a key input for operational decision support systems that help agencies proactively manage the transportation network. This research leverages probe-based traffic speeds, historical weather data, and forecasted weather data to compare the accuracy and computational efficiency of different machine learning and deep learning models on a 6.5-mile corridor in Salt Lake City, Utah. The models are architected to capture both the spatial and temporal dependencies in traffic speed on the roadway network but not require extensive computational resources so that they can be easily scaled. A scalable Partial Least Squares (PLS) Regression model is employed to predict short-term speed using local probe-based traffic data. The training process includes optimizing the number of upstream and downstream roadway links used in the speed prediction for a link, to avoid needing to train on a wider roadway network. Additionally, this study investigates deep learning models using Long Short Term Memory (LSTM) networks, which are well-suited for prediction of time dependent datasets like traffic speed. Performance of the LSTM model is evaluated using different combinations of three input datasets: (1) traffic speed data; (2) historical weather data; and (3) forecasted weather data. The LSTM model that used all three input datasets exhibited slightly better accuracy compared to the other two models, but significantly outperformed them during a major snowstorm event in the test set. The produced models exhibited higher accuracy than other classical/statistical models presented in the literature while remaining geographically scalable.

Ting Li, Massachusetts Institute of Technology
Patrick Meredith-Karam, Massachusetts Institute of Technology
Hui Kong (kongh@mit.edu), University of Minnesota, Twin Cities
Anson Stewart, Massachusetts Institute of Technology
John Attanucci, Massachusetts Institute of Technology (MIT)
Jinhua Zhao, Massachusetts Institute of Technology (MIT)

Estimating passengers’ door-to-door travel time, by walking and public transit, can be complex for large networks with many path alternatives available. Additional complications arise in tap-on only transit systems, where passenger alighting data are not recorded. For one such system, the Chicago Transit Authority, this study compares three estimation methods: assuming optimal path choice given scheduled service, as represented in the General Transit Feed Specification (GTFS); assuming optimal path choice given actually operated bus service, as recorded by Automatic Vehicle Location (AVL) systems; and using inferred path choices based on Automated Fare Collection (AFC) records, as processed with an Origin-Destination-Interchange inference algorithm (ODX). As expected, AVL and ODX-derived travel times are found to be greater than those derived from GTFS, indicating that purely schedule-based travel times underestimate the travel times that are actually available and experienced. These discrepancies additionally manifest in significant spatial variations, raising concerns about potential biases in travel time estimates that do not account for reliability. The findings bring about a more comprehensive understanding of the interactions between transit reliability and passenger behavior in transportation research. Furthermore, these discrepancies suggest areas of future research into the implications of systematic and behavioral assumptions implied by using conventional schedule-based data.

Yi Li, Tongji University
Qing Yu, Tongji University
Weifeng Li, Tongji University
Han Yang, Research Institute of Highway, Ministry of Transport

Traffic congestion is a common problem confronted by a large number of cities. Research has already found that few roads can greatly influence the operation of the road network. Therefore, mitigating traffic congestion should focus on these bottlenecks. However, most traffic management measures only implemented on the bottleneck themselves but failed to consider the interaction between roads. From the perspective of congestion formation and propagation, to support urban traffic management, this paper proposes a methodology to identify spatially compacted key road clusters that highly correlated between travel speed and  traffic volume. A unidirectional-weighted correlation network is constructed and community detection algorithm is applied to partition road segments into key road clusters. Three criteria are used to evaluate clusters including sensitivity, importance and impact coverage. A case study is carried out using the taxi GPS data of Shanghai from May 1 st to 17 th in 2019. The methodology outputs 44 key road clusters. These key road clusters can be classified into 3 types according to their spatial distribution patterns-- along with network skeletons, around transportation hubs, and near bridges. Key road clusters with relatively high importance are distributed along Yan’an Elevated Expressway, South-North Elevated Expressway, and Middle Ring Expressway. The methodology unveils the mechanism of congestion formation and propagation, which can offer significant support for traffic management.

Christopher Day (cmday@iastate.edu), Iowa State University
Toushik Niloy, Iowa State University

Probe vehicle data (PVD) is increasingly used to assess highway performance and for policy decisions such as national travel reliability metrics implemented recently by USDOT. The ubiquity of such data and well-demonstrated power for identifying congestion has led some to utilize distributions of travel time estimates based on PVD to serve as estimates of the actual travel times of vehicles on highways. This study investigates how closely such estimated distributions can resemble a ground truth distribution. A simulation study is carried out to develop a dataset from which both ground truth and estimated distributions are created and compared with each other. The effects of the sample type, aggregation interval, position reporting frequency, and segment definitions are examined. The results show that PVD travel time estimates are sensitive to sample type and less sensitive to other factors. All of the estimated distributions have substantial differences when compared to a ground truth distribution, with the interesting result that they are more different during uncongested conditions than during congestion. Some implications of the findings are presented.

Pengyao Ye, Southwest Jiaotong University
Cheng Li, Southwest Jiaotong University

Travel flows of the urban road are more uncertain and complex than freeways, due to traffic control, stochastic arrivals, and departures at signalized intersections, etc. The travel time distribution of arterial travel flow shows a bimodal or a multimodal curve instead of a unimodal curve, which suggests that arterial travel time can be divided into two components at least, the uninterrupted component and the interrupted component, according to whether vehicles experience signal control delays. In this paper, the travel time adaptive filtering algorithm (AF algorithm) is proposed to consider the intersection signal interruptions and estimate the travel times on a more micro level. Firstly, based on the multimodality of travel time, we modify the generic Gaussian mixture model (GMM) to estimate the travel time distribution, and an expectation-maximization (EM) algorithm is adopted to obtain the free-flow component. Then, an adaptive traffic state discrimination mechanism based on signal control is introduced to estimate the queuing delay. Finally, the signal cycle length at the downstream intersection was conducted as a dynamic time window, and the travel time data is cleaned according to real-time traffic status. The field tests show that this novel method provides a more accurate threshold window, and it can efficiently track the typical variations in arterial travel times while suppressing high-value noise data.

Xiaoxiao Zhang (xz2dk@virginia.edu), University of Virginia
Mo Zhao, Virginia Transportation Research Council
Justice Appiah, Virginia Transportation Research Council
Michael Fontaine, Virginia Transportation Research Council

Travel time reliability quantifies the variability in travel times and has become a critical aspect for evaluating transportation network performance. The empirical travel time cumulative distribution function (CDF) has been used as a tool to preserve inherent information regarding the variability and distribution of travel times. With the advancement of data collection technology, probe vehicle data has been frequently used to measure highway system performance. One challenge with using CDFs when handling large amounts of probe vehicle data is deciding how many different CDFs are necessary to fully characterize experienced travel times. This paper explores statistical methods for clustering CDFs of travel times at segment level into a minimum number of homogeneous clusters that retain all relevant distributional information. Two clustering methods were tested, one based on classic hierarchical clustering and the other used model-based functional data clustering, to find out their performance on clustering distributions using travel time data from Interstate 64 in Virginia. Freeway segments and those within interchange areas were clustered separately. In order to find the proper data format as clustering input, both scaled and original travel times were considered. In addition, non-data driven method based on geometric feature was included for comparison. The results showed that for freeway segments, original travel times using Anderson Darling dissimilarity matrix and Ward’s linkage had the best performance. For interchange segments, model-based clustering provided the best clusters. By clustering segments into homogenous group, the results of this study could improve the efficiency of further travel time reliability modeling groups.