2020

Tung Truong, Illinois Institute of Technology
Ji Zhang, Illinois Institute of Technology
Zongzhi Li, Illinois Institute of Technology
Lu Wang, Illinois Institute of Technology

Transportation investment decision-making is multicriteria in nature, where various types of physical facilities, expectations of stakeholders, and budget constraints are typically involved. In order to obtain decisions that are optimized at the overall system level, a variety of performance criteria in economic, safety, social, and environmental perspectives should be considered to evaluate and prioritize investment alternatives. This paper introduces an iterative multi-objective tradeoff method for optimal capital investment decisions by incorporating Inverse Herfindahl-Hirschman Index (Inverse HHI) into compromise programming (CP) and the ε-constraint method. The Inverse HHI index, which has been widely used for measuring market concentrations and monopolization in economics, is adopted to derive relative importance of multiple non-commensurable performance criteria. Next, CP model helps transform the multi-objective optimization problem to a single-objective problem in the context of minimizing Chebyshev distance to the ideal values. Finally, the ε-constraint method conducts tradeoff analysis between performance criteria effectively and derive best compromised decision outcomes by defining new constraints for each round of iterative process. Six-year data on candidate projects proposed for interstate highway programming in the United States is used for a computational study and methodology validation. The outcomes of model application and iterative tradeoff analysis show that the proposed Inverse HHI-compromise programming in conjunction with ε-constraint method (termed as Inverse HHI-CP ε-constraint method) is capable of generate optimal solutions and effectively conduct budget allocation when additional information is made available. The computation results also show the potential for adopting the proposed method by transportation agencies to develop optimal investment programs.

Gareth Robins, EROAD
Jason Anderson, Portland State University
Salvador Hernandez, Oregon State University

A key component of Vision Zero is identifying and understanding the user attributes and needs that contribute to the incidence of deaths and serious injury in road crashes. As traditional observational methods reach their limits it is inevitable that results should stagnate, and we have to start looking toward new techniques and applications to address the large number of traffic-related crashes that occur each year. In this study, a number of geospatial and machine learning techniques such as DBSCAN, Principal Component Analysis and a modified three-dimensional hausdoorf distance calculation to determine the similarities between clusters of harsh braking locations and crash locations where utilized. The resulting similarities were classified using a machine learning model combining publicly sourced data sets and driver behaviour data from EROAD, a global regulatory telematics company. When applied to the New Zealand road network, the result is a targeted list of specific locations that are susceptible to experience crashes, and a new method for evidence-based interventions that the road controlling authority can use.

Pouria Asadi, University of Rhode Island
Mayrai Gindy, University of Rhode Island
Marco Alvarez, University of Rhode Island
Alireza Asadi, Azad University

Rebar detection in bridge deck Ground Penetrating Radar (GPR) B-scans using Convolutional Neural Networks (CNN) for on-site applications is addressed. The authors investigated accuracy/frame rate trade-off of modern deep learning detection methods for automatic rebar detection in GPR B-scans on ARM platforms. The cost, portability, power consumption and Thermal Design Power (TDP) are advantages of ARM processors over parallel computing platforms, makes them a good options for wide range of applications including robotic and drone based bridge inspection. The authors review three recent meta-architectures (R-FCN, Faster, and SSD, decoupling the choice of meta-architecture from feature extractor so that effect of various feature extractors is investigated to obtain the best option for detecting rebars in B-scans on mobile and embedded systems. State of the art results is obtained for ARM platform based detection task on GPRDETN dataset by implementing rebar detection model using SSD as meta-architecture and Inception v2 as feature extractor. The obtained results indicates that SSD meta-architecture with MobileNet as feature extractor achieves overall mAP of 46 on GPRDETN dataset and remained stable during stress test on ARM platform. SSD meta-architecture with Inception v2 feature extractor outperformed MobileNet detection model by 16.6% with overall mAP of 56 and remained stable by increasing the memory allocated to the processor. A dataset of 520 bridge deck B-scans and 4,085 instances, named GPRDETN, annotated for detection tasks according to Pascal VOC, is created and introduced in this paper for first time

Bingrong Sun, University of Virginia
Lin Gong, University of Virginia
Jisup Shim, Delft University of Technology
Kitae Jang, Korea Advanced Institute of Science and Technology (KAIST)
B. Brian Park, University of Virginia
Hongning Wang, University of Virginia
Jia Hu, Tongji University

Current route guidance systems have simplified assumptions about drivers’ route choice preferences and do not adequately accommodate drivers’ heterogeneous route choice preferences. Main challenges that route guidance systems do not consider drivers’ heterogeneous preferences include: (i) difficulty in acquiring exogenous criteria (e.g., sociodemographic information) that are typically used to differentiate drivers’ preferences; and (ii) difficulty in capturing preference of individuals due to limited preference data. To address these, this paper introduced a human-centric machine learning technique called Multi-Task Linear Classification Model Adaption (MT-LinAdapt) model. It can capture drivers’ common aspects of route choice preferences and yet adapts to each driver’s own preference. In addition, any evolvement of individual drivers’ preferences can be simultaneously integrated to update the captured common preference for further individual drivers’ preference adaptation. MT-LinAdapt was evaluated against two state-of-the-art route choice preference modeling approaches including an aggregate model and an individual model, which are categorized based on the data used for modeling. With a dataset developed from roadside detector records that were collected in Daegu city, South Korea, MT-LinAdapt was compared to existing modeling approaches for its performance when (i) different amounts of data is available, and (ii) when data contains different levels of preference heterogeneity. The evaluation results showed that, compared to existing approaches, MT-LinAdapt achieved up to 28% higher prediction accuracy when limited amount of data with high preference heterogeneity is available; and at least 2% higher prediction accuracy when adequate amount of data with low preference heterogeneity is available.

Zhiyong Cui, University of Washington
Longfei Lin, Beihang University
Ziyuan Pu, University of Washington
Yinhai Wang, University of Washington

Traffic forecasting is a classical task for traffic management and it plays an important role in intelligent transportation systems. However, since traffic data are mostly collected by traffic sensors or probe vehicles, sensor failures and the lack of probe vehicles will inevitably result in missing values in the collected raw data for some specific links in the traffic network. Although missing values can be imputed, existing data imputation methods normally need long-term historical traffic state data. As for short-term traffic forecasting, especially under edge computing scenarios, traffic forecasting models with the capability of handling missing values are needed. To overcome those problems, in this study, we propose a graph Markov network (GMN), which is a new neural network architecture for spatial-temporal data forecasting. Unlike other existing recurrent neural network (RNN)-based models dealing with traffic data as multivariate time series, the GMN handling the traffic state transition process as a graph Markov process. In this way, the proposed GMN incorporates the spatial-temporal relationships among the links in the traffic network. Experiments of traffic forecasting with missing values tested on real-world traffic state datasets show that the proposed GMN achieves state-of-the-art prediction performance in terms of both accuracy and efficiency.

Ruimin Ke, University of Washington
Yifan Zhuang, University of Washington
Ziyuan Pu, University of Washington
Yinhai Wang, University of Washington

Cloud computing has been a main-stream computing service for years. Recently, with the rapid development in urbanization, massive video surveillance data are produced in an unprecedented speed. A traditional solution to deal with the big data would require large amount of computing and storage resources. With the development of Internet of things (IoT), artificial intelligence, and communication technologies, edge computing offers a new solution to the problem by processing the data partially or completely on the edge of a surveillance system. In this study, we investigate the feasibility of using edge computing for smart parking surveillance tasks, which is a key component of Smart City. The system processing pipeline is carefully designed with the consideration of flexibility, real-time surveillance, data transmission, detection accuracy, and system reliability. It enables artificial intelligence at the edge by implementing an enhanced single shot multibox detector (SSD). A few more algorithms are developed either on the edge or the server targeting optimal system efficiency and accuracy. Thorough field tests were conducted in the Angle Lake parking garage for three months. The experimental results are promising that the final detection method achieves over 95% accuracy in real-world scenarios with high efficiency and reliability. The proposed smart parking surveillance system can be a solid foundation for future applications of intelligent transportation systems.

Zhensong Wei, University of California, Riverside
Yu Jiang, University of California, Riverside
Xishun Liao, University of California, Riverside
Xuewei Qi, General Motors Company
Ziran Wang, University of California, Riverside
Guoyuan Wu, University of California, Riverside
Peng Hao, University of California, Riverside
Matthew Barth, University of California, Riverside

This paper presented a deep reinforcement learning method named Double Deep Q-networks to design an end-to-end vision-based adaptive cruise control (ACC) system. A simulation environment of a highway scene was set up in Unity, which is a game engine that provided both physical models of vehicles and feature data for training and testing. Well-designed reward functions associated with the following distance and throttle/brake force were implemented in the reinforcement learning model for both internal combustion engine (ICE) vehicles and electric vehicles (EV) to perform adaptive cruise control. The gap statistics and total energy consumption are evaluated for different vehicle types to explore the relationship between reward functions and powertrain characteristics. Compared with the traditional radar-based ACC systems or human-in-the-loop simulation, the proposed vision-based ACC system can generate either a better gap regulated trajectory or a smoother speed trajectory depending on the preset reward function. The proposed system can be well adaptive to different speed trajectories of the preceding vehicle and operated in real-time.

Zhongxia Xiong, Beihang University School of Transportation Science and Engineering
Ziying Yao, Beihang University
Xinkai Wu, Beihang University
Ming Li, Beihang University

In this paper, a vehicle Re-ID framework which integrates image processing and traffic flow model is developed. First, the CNN network is applied for vehicle detection and tracking and extracting attribute recognition. Particularly, attributes including vehicle color, type, make, and globle feature are extracted to derive a similarity matrix between upstream and downstream vehicles. However, solely using these features could not achive satisfaction matching accuracy. Our testing only shows a moderate accurate of around 75%. To further improve the Re-ID rate, this paper integrates visual information with the well-known IDM car-following mode. In our framework, IDM is first used to estimate the arrival time window for each upstream vehicle; and then with this time window derive a filter matrix which set the similarity as 0 for the matching vehicles outside the time window. Combining similarity matrix and filter matrix, the new developed Re-ID framework improves the matching rate to 96%. Furthermore, the proposed framework can even help identify vehicles that may have changed lanes, overtaken vehicles or driven on a sideroad. Such information is certainly valuable for future research on performance measure, traffic control, and congestion mitigation. Considering the significance of the trajectory data to nowadays traffic control and management and popularity of today’s surveillance cameras, this research certainly will contribute to the improvement of arterial traffic performance measure and efficient control.

Alireza Mostafizi, Oregon State University
Matthew Frantz, Oregon State University
Haizhong Wang, Oregon State University

The cost of congestion in the US is hundreds of billions of dollars annually, and it is estimated that on average, city-dwellers lose thousands of dollars while sitting in a traffic jam for 42 hours every year. Aiming for better mobility and more efficient utilization of the transportation network, emerging connected and autonomous vehicle (CAV) technologies and the resulting communication capabilities can produce more coordinated and efficient routing behavior. Current routing strategies either rely on a centralized control system which can fail in scaling, or employ decentralized schemes that yield sub-optimal coordination and accordingly, poor system performance. This paper introduced a Decentralized Collaborative Time-dependent Shortest Path Algorithm (Dec-CTDSP) with which the CAVs optimize their route according to the communicated mobility messages regarding the location, speed, and the preferred path of the other CAVs within their cluster. We analyzed the impacts of this optimization scheme under various levels of CAV market penetration and communication radius. The results of this study reveal (1) Up to 40% improvements in mean system travel time and speed; (2) Up to 45% increase in travel time and speed prediction reliability; (3) A strong correlation between system travel time and the network usage distribution; and (4) significant improvements in network utilization as a result of Dec-CTDSP. The performance of Dec-CTDSP is further benchmarked against other Dijkstra-based and random algorithms. The findings of this work will help steer further research on the implementation of coordinated and decentralized multiagent routing optimization in the context of connected and autonomous vehicles.

Tingting Zhao, University of Maryland, College Park
Di Zhuang, University of South Florida
Yu Zhang, University of South Florida

The computational burden of optimization-based or simulation-based system performance evaluation, especially for large road networks, limited most of the critical component identification literature to solve only single link failure scenario. In this study, a novel road network critical link combination identification framework is proposed, which is formulated as a learning problem by leveraging network embedding techniques. It is composed of (i) link-level feature extraction, (ii) network-level feature generation, and (iii) learning-based critical link combination identification. Three categories of features (transportation network design features, transportation network flow features, and transportation network topology features ) are defined and extracted from the road networks with multiple links disrupted and taken as link-level features. Adapting from the idea of bag-of-words (BOW) model in natural language processing, the network-level features are generated based on the link-level features to make the proposed model generalizable regarding to various road networks sizes, topologies, and properties. Three different learning techniques (i.e., regression, classification, and ranking) are utilized to determine the critical link combinations. Our extensive experimental evaluations for four transportation networks, various feature combinations, and three learning models, demonstrated and verified that the proposed framework was effective, efficient, scalable, generalizable, and expansible for solving the critical link combination identification problem, with F1-score larger than 0.9 for most cases and 80x-260x faster than the benchmark traffic assignment model.

Yinpu Wang, Southeast University
Leilei Zhu, Highway Bureau of Jiangsu Provincial Transportation Department
Jishun Ou, Southeast University
Zhenbo Lu, Southeast University
Jingxin Xia, Southeast University

As a typical category of automatic vehicle identification (AVI) data, automatic license plate recognition (ALPR) records are usually used to infer vehicle trajectories which play a significant role in many applications of intelligent transportation system (ITS). In practice, the obtained vehicle trajectories are often incomplete due to the low sampling rate resulting from sensor malfunction or complex traffic conditions. Therefore, many researchers investigated various vehicle trajectory reconstruction methods. In recent years, deep learning has achieved great success in many domains because of its great ability to capture complex and hierarchical features and patterns from available data. In this paper, a framework based on deep recurrent neural networks (RNNs) is proposed to accurately reconstruct vehicle trajectories using ALPR data. The framework consists of two key modules. The first is a trip extraction module based on the Isolation Forest algorithm. The second is a vehicle trajectory reconstruction module built on deep long short-term memory (LSTM) neural networks. The ALPR data collected from a realistic network in Kunshan, China was chosen to evaluate the performance of the proposed framework. The test sets were generated by using different sampling rates ranged from 30% to 80%. Experimental results show the framework can achieve 91% accuracy with an 80% sampling rate, and achieve 85% accuracy with a 30% sampling rate, demonstrating its potential ability in ITS applications. Moreover, it is also shown that the reconstructed trajectories can be used to estimate accurate link flows.

Lulu Tan, McGill University
Xudong Wang, McGill University
Luis Miranda-Moreno, McGill University
Aurelie Labbe, HEC Montreal
Lijun Sun, McGill University

The traffic system is a system of spatiotemporal distributions. Spatiotemporal traffic data, which can be regarded as multivariate time series, are often collected using a network of sensors. However, sensor technologies or any other data collection methods are not flawless, as factors ranging from technology malfunction to human error can cause incompleteness in data. Hence, the  missing data problem in the traffic field is often unavoidable. This can be a hindrance to the the performance of data-driven intelligent transportation system (ITS) applications, and other subsequent traffic prediction tasks. Thus, it is essential to develop a reliable imputation method that can help recover missing data as accurately as possible. In this paper, we propose a framework for incorporating the spatial correlation of road network topology, and the temporal dependencies of time-series by building a spatiotemporal regularized factorization model to impute for missing traffic data. Specifically, we use  weighted Laplace matrix and  temporal spline graph as a smoothing approach for retaining and finding the global structure similarity  in  the spatial and temporal dimensions, respectively. We examine the effectiveness of the proposed spatiotemporal regularized factorization model on a traffic volume data set. The result shows the model that is spatiotemporal regularized, achieved the best imputation accuracy compared to the model that is not spatiotemporal regularized. We also discover that compared with spatial regularization, the temporal regularization has the most impact on the imputation result. Overall, this paper demonstrate the importance of incorporating both the spatial and temporal correlations when modelling traffic data.

long chen, University of Glasgow
Piyushimita Thakuriah, Rutgers University
Konstantinos Ampountolas, University of Thessaly

As ride-hailing services become increasingly popular, being able to accurately predict demand for such services can help service operators allocate drivers to customers efficiently, thereby reducing drivers' waiting/idle time, fuel consumption and traffic congestion. In this paper, we propose UberNet, a deep learning convolutional neural network, for short-time predictions of demand for ride-hailing services. Exploiting traditional time series approaches for this problem is difficult due to strong surges and declines in pickups, as well as spatial concentrations of demand, leading to pickup patterns that are  unevenly distributed over time and space. UberNet employs a multivariate framework that utilises a number of temporal and spatial features that have been found to explain demand for ride-hailing services. Specifically, the proposed model includes two sub-networks that aim to encode the source series of various features and decode the predicting series, respectively. To assess the performance and effectiveness of  UberNet to short-term Uber demand predictions, we use six months of Uber data in 2014 and exogenous metadata from the New York City. By comparing the performance of UberNet with several other prediction approaches (ARIMA, PROPHET, and LSTM), we show that the model is highly competitive. The  Ubernet model's performance is better when using economic, social and built environment features to predict Uber pickups,  which suggests that this architecture is more naturally suited when introducing multiple complex social patterns that motivate travel behavior, in real-time demand predictions of ride-hailing.

Zhongxia Xiong, Beihang University School of Transportation Science and Engineering
Ziying Yao, Beihang University
Xinkai Wu, Beihang University
Ming Li, Beihang University

This paper proposes a method for vehicle trajectory reconstruction at single intersection with multiple camerasusing image-Based vehicle detection, tracking and re-identification technique. The proposed method is deivided into a framework for trajectory reconstruction under single camera and reconstruction algorithm for multiple cameras. Firstly, single camera trajectory reconstruction framework contains a real-time vehicle detector for locating vehicles, together with a vehicle tracker for generating single camera trajectories.Using multiple threshold matching for training and an improved classification module, the detector gains a score of 77.29 on UA-DETRAC benchmark with a real-time speed of 27.78 fps. Moreover, the proposed vehicle re-identification (ReID) model shows a competitive score of 59.77 on VeRi dataset. ReID model is then integrated with a current leading object tracker and obtains a significant improvement also on UA-DETRAC. Reconstructed trajectories in each camera are mapped to a unified coordinate system thereafter. Then, for trajectory reconstruction with multiple cameras, the collected trajectories are firstly separated by traffic signal. In each duration of the signal, for a vehicle driving into the intersection must drive in a restricted area due to the given phase. With these temporal and spatial constraints, Kalman filter and vehicle re-identification features are integrated for trajectory prediction and updating. Finally, overlapping trajectories are merged and the gaps of area that all cameras fail to cover is filled. The results of successfully reconstructed complete trajectories are verified on Cityflow dataset, presenting a verification score of 93.23.

Zemian Ke, Southeast University
Zhibin Li, Southeast University
Pan Liu, Southeast University
Yong Liu, Nanjing University

The primary objective of this study was to propose a framework which incorporates the deep reinforcement learning algorithm into the variable speed limit (VSL) control strategies to reduce total time spent (TTS) at freeway bottlenecks . A double deep Q network (DDQN)-based VSL control strategy, which allows using continuous traffic state representation and neural networks for Q-value estimation, was proposed for the system performance optimization . The action was speed limit, and the reward was aimed at reducing TTS . An off-line training procedure was proposed for the DDQN agent based on iterations with the simulation of the cell transmission model (CTM) to determine the optimal speed limits for various traffic states to reduce congestion . Two typical scenarios including the stable demand and the fluctuating demand were considered to fully evaluate the online performances of the proposed strategy . The results showed that the proposed DDQN-based VSL control strategy reduced the TTS significantly both in the stable and the fluctuating demand scenarios . The effects of the DDQN-based VSL strategy were compared with those of the Q-learning (QL)-based VSL strategy proposed previously . The results showed that the proposed DDQN-based VSL strategy performed better than QL-based VSL strategy in reducing travel time .

Shenghong Dai, University of Wisconsin, Madison
Changyan Fan, Southeast University
Mofeng Yang
Zhibin Li, Southeast University
Yuxuan Hou, Southeast University

Control strategy of ramp metering (RM) decides how ramp flow merges into freeway mainline according to traffic conditions, which significantly affects traffic operation. In this paper, we proposed a deep reinforcement learning (DRL)-based RM control framework with the image-input convolutional neural network (CNN). The proposed framework established the real-time and full-space traffic state analysis based on direct image-input from the study area, which overcame the approximate traffic states estimated from traditional fixed location loop detectors. Deep Q Network (DQN) with prioritized experience replay as one of DRL algorithms was integrated into the RM control tasks in order to reduce total travel time on freeways. The open-source simulation software Simulation of Urban Mobility (SUMO) was used to simulate a typical freeway weaving bottleneck, train the DQN agent and evaluate RM strategies. The results showed that the proposed Image-DQN-based RM strategy was able to respond proactively to various traffic states perceived by processed images, obtain the optimal control strategy in short training time, and take immediate actions to prevent traffic breakdowns. With the proposed Image-DQN-based RM, the total travel time was reduced by 43.48% and 42.24% with 15-s and 30-s traffic control cycle. The performances of various RM were also compared. The results showed that the proposed Image-DQN-based RM outperformed the traditional fixed-time RM, the feedback-based RM strategy, and also the DQN-based RM control strategies with loop detector data as input in mitigating congestions and reducing travel time on freeways.

Zhiyan Yi, University of Utah
Xiaoyue Cathy Liu, University of Utah
Ran Wei, University of California, Riverside
Tony Grubesic, University of Texas, Austin

Snowplow trucks serve a crucial role in winter maintenance activities by removing, loading and disposing snow. An effective performance monitoring and analysis process can assist transportation agencies in effectively managing the snowplow trucks and maintaining normal functioning of roadways. Previous literature suggests that most snowplow truck performance analysis is done through cost-benefit analysis at the macro-level to determine the optimal life cycle for the entire truck fleet. However, the proposed optimal life cycle could lead to waste of resources, and may incur bias due to the ignorance of performance variations resulting from endogenous and exogenous features. More importantly, it fails to identify the contributable factors to performance deterioration. With the proliferation of data in recent years, the aforementioned concerns can be addressed through predictive machine learning techniques in a data-driven fashion. In this study, we apply machine learning techniques, including the random forest (RF) algorithm and a support vector machine (SVM) to predict the performance of snowplow trucks. Using the snowplow truck fleet managed by the Utah Department of Transportation (UDOT), both models are implemented and it is demonstrated that RF outperforms linear SVM with regard to prediction accuracy. Further, a feature importance analysis can assist transportation agencies to improve truck replacement strategy by identifying crucial factors for their performance. Lastly, a sample application of the developed prediction model suggests the threshold of work intensity for preventing rapid deterioration of trucks’ performance under various working environments.

Raed Hasan, Western Michigan University
Hafez Irshaid, General Motors Company
Sangwoo Lee, Western Michigan University
Jun-Seok Oh, Western Michigan University

Transportation Mode Detection (TMD) is important in planning new transportation projects as well as improving existing ones. Therefore, this study aims to develop predictive modes of transportation through the use of smartphone data and smartwatches, as well as the use of machine learning techniques. To achieve the objective of this study, and as part of it, a review of the studies related to the use of algorithms to predict transportation modes was prepared. Besides, on the practical side, the platform of Physical Activity for Smart Travel (PASTA) has been developed. Two groups of participants were recruited in the states of Michigan and Texas to obtain the required data for the study. Daily activities have been classified into activities related to transportation (trips) and non-related to transportation (home, work, shopping, etc.), then focusing exclusively on transportation activities to determine their modes. In this study, four machine learning methods were used for prediction (Extreme Gradient Boosting, Random Forest, Support Vector Machine, and Artificial Neural Network), as well as using the data of physical activities as a new feature not used in previous studies. The accuracy of the methods of transportation mode prediction was compared through the training and testing phases and the results of the predictions were compared with the activities verified by the participants. The results showed that the method of Random Forest worked better than other methods. This study provides the appropriate tools for decision makers to help them understand the travel behaviors of people.

Zhengwei Bai, Beijing Jiaotong University
Peng Hao, University of California, Riverside
Matthew Barth, University of California, Riverside

Taking advantage of both vehicle-to-everything (V2X) and autonomous driving technology, connected and automated vehicle (CAV) emerges as one of the transformative solutions to the current transportation problems. The traffic around signalized intersection is one good scenario to fuse connectivity and automation for improving vehicle mobility and energy efficiency, as it need on-board sensors to perceive surrounding vehicles and V2X communication to interact with traffic signal timing and non-surrounding vehicles. However, it is a challenge to effectively integrate sensor-based information and V2X message in the mixed connected traffic considering the complexity and uncertainty in the traffic system. In this study, we proposed a hybrid reinforcement learning (HRL) framework which combines the rule-based strategy and the deep reinforcement learning (deep RL) to support connected eco-driving at signalized intersections. Vision-perceptive methods are integrated with vehicle-to-infrastructure (V2I) communication achieve high mobility and energy efficiency in mixed connected traffic. The HRL framework is composed of three main parts: a rule-based driving manager that manages the collaboration between rule strategy and RL policy; a multi-stream neural network that extracts the hidden features of vision and V2I information; and a deep RL-based policy network that generate both longitudinal and latitudinal eco-driving actions. In order to evaluate the method, we developed a Unity-based simulator and designed an intersection scenario which includes mixed human-driven vehicles. Moreover, several baselines are implemented to compare with the designed method and numerical experiments are conducted to test the performance of the HRL model. The experiments show that the HRL method can save 1.13%-7.58% travel time and reduce 12.25%-47.13% energy consumption comparing with several baselines.

Michael Brogan, Reflective Measurement Systems
James Mahon, Reflective Measurement Systems

This paper describes a Convolutional Neural Network (CNN) developed for the purpose of identifying and classifying images of retroreflective central lane markings. These may include images that would be difficult to identify using classical machine vision techniques due to degradation. Inference can be performed at speeds up to 100 frames per second. Once identified and classified using the CNN, the percentage of retroreflective material present on the symbol may be expressed as a percentage of the area of non-retroreflective material. The techniques were developed from data acquired by a third-generation mobile retroreflectometer during the national Irish survey from 2018. The development took the form of four phases: 1) data acquisition and manual classification; 2) CNN development and classification; 3) integration of the CNN inference with the existing survey processing software platform, and; 4) retroreflective coverage metrics. Classification accuracy of >98% was achieved using the CNN, allowing markings to be inferred as one of four classes, accounting for 93.86% of the detected central lane markings. Following classification, three metrics were calculated to determine the functionality of the marking: non-retroreflective area, retroreflective area, and the percentage of retroreflective material present compared to non-retroreflective material.

Ouafa Benkraouda, New York University, Abu Dhabi
Bilal Thonnam Thodi, New York University, Abu Dhabi
Hwasoo Yeo, Korea Advanced Institute of Science and Technology (KAIST)
Monica Menendez, New York University, Abu Dhabi
Saif Jabari, New York University, Abu Dhabi

We propose a statistical learning-based traffic speed estimation method that uses sparse vehicle trajectory information. Using a convolutional encoder-decoder based architecture, we show that a well trained neural network can learn spatio-temporal traffic speed dynamics from time-space diagrams. We demonstrate this for a homogeneous road section using simulated vehicle trajectories, and then validate it using real-world data from NGSIM. Our results show that with probe vehicle penetration levels as low as 5%, the proposed estimation method can provide a sound reconstruction of macroscopic traffic speeds and reproduce realistic shockwave patterns, implying applicability in a variety of traffic conditions. We further discuss the model's reconstruction mechanisms and confirm its ability to differentiate various traffic behaviors such as congested and free flow traffic states, transition dynamics, and shockwave propagation.

Jonathan Waddell, Wayne State University
Stephen Remias, Wayne State University
Jacob Holden, National Renewable Energy Laboratory (NREL)
Henry Fournier, Wayne State University
Mark Brown, Wayne State University

Vehicle trajectory data from cell phones, GPS devices, on-board units, and in-vehicle telematics are rapidly becoming a valuable data source for the measurement of traffic operations. Currently, transportation practitioners use metrics such as travel time, delay, queues, or arrivals on green as objective functions to optimize transportation systems. In the future, energy or fuel consumption may be a viable alternative to those traditional metrics. The objective of this study was to establish the feasibility of recurrent neural networks for prediction of mass air flow rates (MAF), which are commonly used to derive fuel consumption rates. On-board diagnostic (OBD) data was collected from four different vehicles while driving four arterial corridors in Detroit, MI. Vehicle trajectory data was used as an input into a recurrent neural network and the results were compared with measured values from OBD data and a vehicle powertrain based model. The r-squared value when comparing the predicted MAF to measured MAF for the four vehicle types ranged from 0.84 to 0.89. The mean percent error of the fuel predictions for each trip were comparable to a powertrain based physics model, 10.97% and 12.12% respectively. The results show that the recurrent neural network output is comparable and could potentially be used to scale vehicular energy use statistics nationwide.

Zhenyu Wang, Old Dominion University
Hong Yang, Old Dominion University
Mecit Cetin, Old Dominion University
Zhitong Huang, Leidos, Inc.
Sudhakar Nallamothu, Leidos, Inc.
Peter Huang, Federal Highway Administration (FHWA)

Traffic incident detection is a crucial task in traffic management centers (TMCs) that typically manage many highways with limited staff. An effective automatic incident detection approach that can report abnormal events timely and accurately will benefit TMCs in optimizing the use of limited incident response and management resources.   During the past decades, significant efforts have been made by researchers towards the development of loop-based automated approaches. Nevertheless, many developed approaches have shown limited success in the field. This is largely because of the long detection time (waiting for overwhelmed upstream detection stations; meanwhile, downstream stations show light traffic volume), and the concerns about the costly false alarms (e.g., dispatching response teams to nonincident cases).   With the advancements in artificial intelligence (AI) technology, there is an opportunity to use AI to significantly improve incident detection practice. This paper proposes an AI-based incident detection framework that can leverage large-scale sensor data along with advanced learning algorithms to enhance the predictive performance. It investigates the generic algorithmic problems when designing a detection approach and places more emphasis on the architecture of the AI-based detection framework with the inclusion of learning and evolving capabilities. The proposed framework has been assessed by case studies. When compared against the conventional approaches, the test results indicate that the proposed AI-based framework has advantages in achieving higher detection rates, less false alarm rates, and shorter time to detect the incidents in various experimented scenarios. Some extensions of the proposed framework are also discussed.

Zirui Raymond Huang, University of Arizona
Ali Arian, Metropia Inc.
Yuqiu Yuan, University of Arizona
Yi-Chang Chiu, University of Arizona

An increasingly emphasized research area is the forecast of short-term traffic conditions for nonrecurring traffic dynamics caused by random highway incidents such as crashes or roadway closures. This research proposes a prediction framework which focuses on training a machine learning (ML) model to predict the speed heatmap associated with incidents. Heatmaps contain ideal information that depicts the spatiotemporal characteristics of incidentinduced impacts and are suitable objects for ML models to understand and predict. Because of the sparsity of incident data in the real world, we proposed a simulation approach to speed up the model training. The Conditional Deep Convolutional Generative Adversarial Nets (C-DCGAN), is employed to predict the speed heatmap and the mesoscopic Dynamic Traffic Assignment (DTA) model DynusT was used to generate a large number of training data. The evaluation shows that the proposed model captures both the tonal and spatial distribution of pixel values at 80.19% similarity between the prediction and actual heatmaps. To the best of our knowledge, this is one of the first attempts in the literature to train ML to predict heatmap representation of incident-induced spatiotemporal impact, and speeding up the training via simulation.

Anna Philips, University of Texas, Arlington
Farah Naz, University of Texas, Arlington
Kate Hyun, University of Texas, Arlington
Vivek Patel, University of Texas, Arlington
Gordon Zhang, Georgia Institute of Technology (Georgia Tech)
Won Hwa Kim, University of Texas, Arlington

Stakeholders and transportation planners often use users’ feedback to assess transit services including ride hailing platforms and reflect them for future plans. Interestingly, social network services (SNS) provide such information in a large set of text by individuals exchanging event base attitude and sentiment. This information is very useful, however, these data are often unorganized and it is intractable to process this extremely large set of text data by human effort whose size is continuously increasing. In this regime, we collected ride hailing service relevant text data from Twitter and created a database, and developed a novel Deep Learning (DL) framework that process and classify sentences that will automatically categorize the texts uploaded by service users according to transportation service specific criteria’s. Our model uses multiple kernels for convolution to capture local context among neighboring words in texts and is simplified by summarizing parameters in traditional models using a kernel function. Using our DL model, we trained a classifier that identifies 1) to which transit service a text corresponds (e.g., reliability, mobility and cleanness), and 2) which sentiment the text contains (i.e., positive vs. negative). Its prediction performance is comparable to state-of-the-art DL methods but our model converges much faster during training which means it trains much more efficiently. We expect that our framework will provide feedbacks for policy makers who explore communication and information technology to create strategies to improve system efficiency and transit ridership.

Nassim Sohaee, University of Texas, Dallas
Farzad Karami, University of Texas, Dallas
Shahram Bohluli, C&M Associates, Inc.
Chao Huang, C&M Associates, Inc.

Traffic forecasting plays a crucial role in the effective operation of managed lanes, as traffic demand and revenue are relatively volatile given parallel competition from adjacent toll-free general purpose lanes. This paper proposes a deep learning framework to forecast short-term traffic volumes and speeds on managed lanes. A network of convolutional neural networks (CNN) was used to detect spatial features. Volume and speed were converted into heatmaps feeding into the CNN layers and temporal relationships were detected by a recurrent neural network (RNN) layer. A dense layer was used for the final prediction. Six months of historical volume and speed data on the I-580 Express Lanes in California, United States were utilized in this case study. Computational results confirm the effectiveness of the proposed data-driven deep learning framework in forecasting short-term traffic volumes and speeds on managed lanes.

Pinchao Zhang, Carnegie Mellon University
Fei Fang, Carnegie Mellon University
Sean Qian, Carnegie Mellon University

In this study we propose a reinforcement learning framework to dynamically route system-level traffic with the consideration of uncertain demand.  Reinforcement learning has the ability of capturing useful information from the interaction with a complex system, and learn a good policy to take actions to improve system performance. While existing methods of traffic routing have certain shortcomings, the proposed  framework overcomes those limitations and shows promising performance. On a small network which the optimal control policy has been analytically solved, experiments show that the proposed framework can find the optimal policy. On the same network, we show that when the variance of the demand becomes relatively large, or there are more vehicles would follow our suggestions, our method has better performance. It also shows better performance on a medium size network than baseline algorithms.

Xilei Zhao, University of Florida
Zhengze Zhou, Cornell University
Xiang Yan, University of Michigan, Ann Arbor
Pascal Van Hentenryck, Georgia Institute of Technology (Georgia Tech)

Machine learning has proved to be very successful for making predictions in travel behavior modeling. However, most machine-learning models have complex model structures and offer little or no explanation as to how they arrive at these predictions. Interpretations about travel behavior models are essential for decision makers to understand travelers' preferences and plan policy interventions accordingly. Therefore, this paper proposes to apply and extend the model distillation approach, a model-agnostic machine-learning interpretation method, to explain how a black-box travel mode choice model makes predictions for the entire population and subpopulations of interest. Model distillation aims at compressing knowledge from a complex model (teacher) into an understandable and interpretable model (student). In particular, the paper integrates model distillation with market segmentation to generate more insights by accounting for heterogeneity. Furthermore, the paper provides a comprehensive comparison of student models with the benchmark model (decision tree) and the teacher model (gradient boosting trees) to quantify the fidelity and accuracy of the students' interpretations.

Sara Zahedian, University of Maryland, College Park
Przemyslaw Sekula, University of Maryland, College Park
Amir Nohekhan, University of Maryland, College Park
Zachary Vander Laan, University of Maryland, College Park

Traffic volumes are an essential input to many highway planning and design models; however, collecting this data for all road network segments is not practical nor cost-effective. Accordingly, transportation agencies must find ways to leverage limited ground truth volume data to obtain reasonable estimates at scale on the statewide network. This paper aims to investigate the impact of selecting a subset of available Automatic Traffic Recorders (ATRs) (i.e., the ground truth volume data source) and incorporating their data as explanatory variables into a previously developed machine learning regression model for estimating hourly traffic volumes. The study introduces a handful of strategies for selecting this subset of ATRs and walks through the process of selecting them and training models with their data as additional inputs using the New Hampshire road network as a case study. The results reveal that the overall performance of the Artificial Neural Network (ANN) machine learning model improves with the additional inputs of selected ATRs. However, this improvement is more significant when the ATRs are selected based on their spatial distribution over the TMC network. For instance, selecting 8 ATR stations according to the TMC coverage-based strategy and training the ANN with their inputs leads to 19.39% average relative increase in R 2 and average relative reductions of 35.39% and 13.44% in MAPE and EMFR, respectively. The results achieved by this study can be further expanded to create a practical strategy for optimizing the number and location of ATRs through states’ transportation networks.

Shuo Wang, NVIDIA Corporation
Subhadipto Poddar, Iowa State University
Pranamesh Chakraborty, Iowa State University
Anuj Sharma, ETALYC Inc
Skylar Knickerbocker, Iowa State University
Neal Hawkins, Iowa State University

Traffic speed prediction has been a key component for Intelligent Transportation Systems (ITS) over the past few years. Several techniques have been applied to predict traffic speed – the most recent one being deep learning methods. The following paper proposes using Convolutional Neural Network (CNN) to predict the traffic speed profile along a roadway segment over a long time span (for example a full day) based on historic traffic speeds and predicted weather conditions. The model uses CNN layers to effectively extract features from the spatial-temporal structure of the input data and predicts traffic speed at the route level. A fully convolutional structure was used so that the network would be independent on the spatial-temporal dimension of the input data and the pre-trained model can be transferred on traffic speed prediction at difference scales. The proposed fully convolutional deep network was optimized by using both 3-D CNN layers and one-by-one CNN layers. The model was tested on three test days (two normal weekdays and a snowy day) and compared with another CNN model without weather as an input and a random forest regression. The CNN model, without weather, showed marginal changes given a class imbalance, however, both the deep models showed improved performance as compared to the random forest regression. The proposed long-term prediction model is expected to assist traffic engineering and planning practitioners, road users, and freight logistics in planning and altering their movements according to predicted conditions.

Heng Wei, University of Cincinnati
Hedayat Abrishami, University of Cincinnati
Zhuo Yao, University of Cincinnati Medical Center

This paper presents an image processing technique with Gaussian Mixture Models (GMM) for extracting two vehicle features, vehicle length and number of axles in order to classify the vehicles from video, based on Federal Highway Administration (FHWA)’s recommended vehicle classification scheme. In the proposed synthetic framework, four steps are detailed to guide the development of the corresponding algorithms: i.e., object detection, tire region extraction, tire extraction in its tire region, and vehicle classification. There are two stages regarding the vehicle classification. The first stage is the general classification that basically classifies vehicles into 4 categories or bins based on the vehicle length (i.e., 4-Bin length-based vehicle classification). The second stage is the axle-based vehicle classification that classifies vehicles in more detailed classes of vehicles such as car, van, buses, based on the number of axles. The developed models and associated algorithms are tested with the sample video data obtained on a segment of I-275 in the Cincinnati area, Ohio. The evaluation results show a better 4-Bin length–based classification than the axle-based group classification.  There may be two reasons. First, when a vehicle gets misclassified in 4-Bin classification, it will definitely be misclassified in axle-based group classification. The error of the 4-Bin classification will propagate to the axle-based group classification. Second, there may be some noises in the process of finding the tires and number of tires. The testing results provides solid basis for integrating the developed framework and algorithms into a permanent traffic monitoring station.

Juyuan Yin, Tongji University
Xuejian Chen, Tongji University
Jian Sun, Tongji University

Vehicle trajectory reconstruction for urban arterial can provide both full-view and detailed traffic flow information which are meaningful for supporting estimating traffic state, emission, fuel consumption, and optimizing traffic signal control. With the continuous development of Connected and Autonomous Vehicle (CAV) technology, a new data environment in which CAVs provide their own and other near vehicles’ high-frequency trajectories using in-vehicle sensors and GPS devices, has attracted researchers to use such data for vehicle trajectory reconstruction. However, the major challenge for CAV data to be further utilized and explored comes from the low penetration rate of CAVs currently and even in the near future. In this study, Arterial Vehicle Trajectory Reconstruction Method (AVTRM) under low-penetration CAV environment is proposed. First, Shockwave Model is applied to estimate the queuing part of undetected vehicles’ trajectories by using CAVs’ and detected vehicles’ trajectories. Then, Inverse Car-Following Model is proposed based on Car-Following Model and the two models are used to estimate the moving part of individual vehicle trajectory from upstream and downstream, respectively. Last, Particle Filter is applied to fuse two types of estimated trajectories of individual vehicle, i.e., upstream-based and downstream-based. One case study was conducted at a field arterial, i.e., Huanggang Road in Shenzhen, China, to evaluate the proposed AVTRM. The results show that the queue location error and time error of reconstructed trajectories were only 1.6 m and 2.7 s, respectively, and compared with a classic vehicle trajectory reconstruction method, the two indices were dropped by 34.79% and 4.76%, respectively.

Peng Jin, University of Missouri
Xiaofan Shu, University of Missouri, Columbia
Vishal Mandal, University of Missouri
Yaw Adu-Gyamfi, University of Missouri, Columbia

Accurate and timely detection of traffic anomalies–such as accidents, congestion and stranded vehicles–can potentially reduce secondary crashes, improve traffic mobility and ultimately lead to quicker dispatch of response-emergency vehicles. These improvements can also help save the lives of motorists involved in accidents. The use of infrastructure-mounted sensors such as CCTV cameras  for  traffic  anomaly  detection  and  verification  is  gaining  popularity  among  most  state agencies.  Nonetheless,  they  are  driven  by  manual  procedures,  making the  process  tedious  and inefficient.  The  goal  of  this  paper  is  to  develop  an  automatic traffic  anomaly  detection  system based on video data. We first implement a convolutional neural network model based on YOLO, a  popular  deep  learning framework  for  object  detection  and  classification.  The  resulting recognition model, together with a multiobject tracking model(based on intersection over union –IOU)is  used  to  search for and  analyze  different  traffic  scenes  for  anomalies.  Post-processing techniques such as bounding box suppression and adaptive thresholding are used to reduce false alarm  rates  and  improve  the  robustness  of  the  methodology  developed.  At  each stage of  our developments, a comparative analysis is conducted to evaluate the strengths and limitations of the approach.The proposed methodology is evaluated based on F1 and S3 performance metrics. The results show the potential of adopting this technology for large scale traffic surveillance.

Ruimin Ke, University of Washington
Wan Li, Oak Ridge National Laboratory
Zhiyong Cui, University of Washington
Yinhai Wang, University of Washington

Traffic speed prediction is a critically important component of intelligent transportation systems (ITS). Recently, with the rapid development of deep learning and transportation data science, a growing body of new traffic speed prediction models have been designed, which achieved high accuracy and large-scale prediction. However, existing studies have two major limitations. First, they predict aggregated traffic speed rather than lane-level traffic speed; second, most studies ignore the impact of other traffic flow parameters in speed prediction. To address these issues, we propose a two-stream multi-channel convolutional neural network (TM-CNN) model for multi-lane traffic speed prediction considering traffic volume impact. In this model, we first introduce a new data conversion method that converts raw traffic speed data and volume data into spatial-temporal multi-channel matrices. Then we carefully design a two-stream deep neural network to effectively learn the features and correlations between individual lanes, in the spatial-temporal dimensions, and between speed and volume. Accordingly, a new loss function that considers the volume impact in speed prediction is developed. A case study using one-year data validates the TM-CNN model and demonstrates its superiority. This paper contributes to two research areas: (1) traffic speed prediction, and (2) multi-lane traffic flow study.

Bowen Cai, Tongji University
Xuxin Zhang, Tongji University
Xuesong Wang, Tongji University

Driving anger detection is an important topic in traffic safety analysis. Driving anger is getting more and more prevalent and serious in everyday traffic incidents. Many deaths and injuries are related to road rage. Aggressive driving behaviors create kinds of roadway traffic safety hazards. Mitigating potential risk caused by road rage is indispensable to increase overall level of traffic safety. This paper puts forward an integrated computer vision model composed of patter recognition and 14 layers very deep convolutional neural network to recognize driver anger emotion and classify angry driving from natural driving status. Histogram of gradients method was applied to locate faces. Multidimensional analysis in calculating geometric pattern of facial appearance was used to classify driving anger from natural status. To mitigate false call problem, threshold of classifying images into natural was tuned up and images that were classified into anger were all sent to deep learning model to recheck. Driver anger detection algorithm with overall accuracy rate of 86.7% was achieved in the simulator environment. Test scenario was established and Tongji University 8 Degree of Freedom driving simulator was used to collect data from 30 recruited drivers.

Abdullah Shabarek, New Jersey Institute of Technology
Steven Chien, New Jersey Institute of Technology
Soubhi Hadri, Microsoft Corp

The introduction of deep learning and big data analysis may significantly elevate the performance of traffic speed prediction under various weather conditions. Adverse weather may cause mobility and safety issues on roadways because of varying traffic speed and/or poor visibility. Most previous modeling approaches did not consider temporal and spatial data heterogeneity, such as weather conditions. This paper presents a deep learning framework, consisting of recurrent deep learning models, to predict traffic speed under rain, fog, and snow conditions, considering prevailing traffic speed, wind speed, the ratio of volume/capacity, wind bearing, precipitation intensity, and visibility. To ensure the accuracy of speed prediction, different models are assessed under various weather and traffic conditions. The results indicate that the Recurrent Convolutional Neural Network (RCNN) model outperforms others, which can yield an average Root Mean Square Error (RMSE) of 5.5 mph. Considering real-time feeds of weather conditions on a 15-minute basis, a visualization system is developed for displaying predicted traffic speeds on New Jersey freeways. This application is very useful for predicting hot spot congestion segments under normal and adverse weather scenarios.

Weiyi Zhou, University of Maryland
Mofeng Yang
Minha Lee
Lei Zhang, University of Maryland, College Park

To increase traffic mobility and safety, several types of active traffic management (ATM) strategies, such as variable speed limit (VSL), ramp metering (RM), dynamic message signs (DMS), and hard shoulder running (HSR), are adopted in many countries. While all kinds of ATM strategies show promise in releasing traffic congestion, many studies indicate that stand-alone strategies have very limited capability. To remedy the defects of stand-alone strategies, coordinated ATM stategies have caught researchers’ attention and different combinations have been studied. In this paper, we proposed a coordinated VSL and HSR control strategy based on a reinforcement learning (RL) technique, Q-learning (QL). The proposed strategy bridges up a direct connection between the traffic flow data and the ATM control strategies via intensive self-learning processes, thus reduces the need for human knowledge. A typical congested interstate highway, I-270 in Maryland, the United States is selected as the study area to evaluate the proposed strategy. A dynamic traffic assignment (DTA) simulation model is introduced to calibrate the network with real world data and was used evaluate the regional impact of the proposed algorithm. Simulation results indicate that the proposed coordinated control could reduce corridor travel time up to 27%. We also compare the performance of various control strategies. The results suggest that the proposed strategy outperforms the stand-alone control strategies and traditional feedback-based VSL strategy in mitigating congestions and reducing travel time on freeway corridor.

Maged Gouda, University of Alberta
Karim El-Basyouny, University of Alberta
Alexander Epp, University of Alberta

The application of point-based deep learning on LiDAR data is promising for timely, accurate sign detection. This study proposes a novel approach that leverages the use of local geometric features in the training process. Further, a practical methodology is proposed to extract and prepare training datasets. Finally, a series of validation tests were run to investigate the accuracy. The proposed approach proves that with adjustments, the PointNet++ neural network architecture can achieve remarkable results on large metric scale scenes for extracting signs from LiDAR point clouds. Models with different combinations of intensity, roughness, z-gradient, and point densities were trained using labelled data from seven 4 km highway segments in Alberta, Canada. The models’ performance was tested on three highway segments using the precision, recall, and F1-score metrics. Further, the statistical significance of the difference between models was evaluated using permutation tests. Overall, the intensity and z-gradient model, with 4,096 cube points, significantly outperformed all other models in terms of precision, recall, and F1-score at the 95% confidence level on all segments. The results showed improved performance in accuracy and processing times compared to most studies on sign detection from point clouds. The overall per sign detection performance on a 12 km segment shows 99.2% recall and 98% F1-score. The intensity with roughness combination yielded lower performance compared to intensity-only and intensity-z-gradient models. Overall, including z -gradient significantly increased precision, recall, and F1-score, by 9%, 4.9%, and 7.1%, respectively, allowing the model to yield remarkable performance on outdoor scene recognition.

Pangwei Wang, North China University of Technology
Hui Deng, North China University of Technology
Li Wang, North China University of Technology
Mingfang Zhang, North China University of Technology

In recent decades, intelligent connected vehicles can take advantage of the vehicle-to-vehicle/vehicle-to-infrastructure(V2V/V2I) communication technology. Through platoon control based on V2V/V2I, it can achieve group consensus and improve running safety and road capacity effectively. Due to the wireless communication system affected by the traffic environment time-varying limitations (time-delay, packet-dropout, interruption, etc.), the platoon control system has some difficulties in practical applications. Considering the characteristics of V2V communication technology, firstly, the predecessor-leader following was selected as the basic platoon topology and the calculation model of the desired vehicle spacing was established. Secondly, combining with the characteristics of multi-input and multi-constraints in the actual traffic environment, the abnormal conditions in the communication process was considered. Based on MPC, the platoon control method of connected vehicle was established and the constraint condition of string stability was analyzed. Finally, based on Prescan/Matlab, a hardware in loop simulation platform for connected vehicle platoon was designed. The platoon control method was tested in six traffic scenarios including constant speed, variable acceleration, normal and abnormal communication. The final tested results showed that according to different communication environments, the platoon topology can be switched in real time through the platoon control method established in this paper, meanwhile ensuring the string stability and the consistent trend of vehicle spacing, speed and acceleration. The proposed platoon control method for intelligent connected vehicles is suitable for different road environments and communication environments, which provides theoretical basis and technical implementation method for the future application of intelligent connected vehicles.

Samuel Hislop-Lynch, University of Queensland
SangHyung Ahn, University of Queensland
Jiwon Kim, University of Queensland

Convolutional Neural Network (CNN) based localization and classification algorithms are a promising means of collecting comprehensive transport data for a broad range of applications. With this technology, it is possible to detect the presence of vehicles in still or video frames, localize the positions of the vehicles within the frame and identify the class of a detected vehicle (e.g. car, bus, truck, etc.). This technology can also be combined with drone-based video data collection to gain insight into problems which would be difficult to investigate using conventional data collection technology. However, little work has been done to quantify the performance of these algorithms on transport-specific tasks. In order for transport researchers and practitioners to adopt this technology and confidently make use of the data, the capabilities and limitations of these algorithms must be understood. To this end, an empirical investigation was undertaken which quantified the performance of a number of existing CNN algorithm implementations, trained on the popular Common Objects in Context (COCO) dataset. A consumer-grade drone was used to capture aerial footage from a publicly accessible carpark, and the impact of four key variables (known to affect the performance of localization and classification algorithms) was quantified. The variables examined were 1) object size (in square pixels), 2) object pose, 3) object orientation and 4) the presence of shadows. Detections obtained from the CNN-based localization algorithms were compared with hand-generated ground truths. An analysis was conducted to determine the impact of each key variable on the localization and classification performance.

Zhiyong Cui, University of Washington
Mingjian Fu, University of Washington
Meixin Zhu, University of Washington
Xuegang Ban, University of Washington
Yinhai Wang, University of Washington

The advancement of new smart traffic sensing, mobile communication, and artificial intelligence technologies has greatly stimulated the growth of transportation data. The increase of computation power enabled by advanced hardware and the rise of artificial intelligence (AI) technologies provides great opportunities to comprehensively utilize the transportation big data. Transportation domain knowledge is beneficial for designing AI models and solving transportation problems. However, because most AI algorithms were not originally designed for transportation problems, using big data and AI technologies to solve transportation problems is facing challenges. Since key hyper-parameters are missed in some proposed AI models, many proposed AI-based methods can hardly be accurately re-implemented. Further, in most of the AI-based transportation research studies, there is no uniform dataset to evaluate the proposed models. In this study, to overcome the challenges mentioned earlier, we propose a transportation AI platform with widely accepted datasets, provide well-established models, and use uniform training and testing procedures to assist the evaluation of emerging novel methodologies. We design a novel architecture for platform to enhance the efficiency of the transportation data processing, management, and communication and increase the computational power of the platform. Traffic forecasting involving high-dimensional spatiotemporal data is a good applicable scenario to utilize novel deep learning models to solve complicated transportation problems. The developed transportation AI platform is capable of evaluating the traffic prediction performance of various implemented models by comparing and visualizing the prediction results tested on multiple real-world network-wide traffic state data sets.

Saleh Mousa, Texas Department of Transportation
Sherif Ishak, Old Dominion University
Ragab Mousa, Cairo University
Julius Codjoe, Louisiana Department of Transportation and Development
Mohammed Elhenawy, Queensland University of Technology

Eco-Approach and Departure is a complex control problem wherein a driver’s actions are guided over a period of time or distance so as to optimize fuel consumption. Reinforcement Learning (RL) is a machine learning paradigm that mimics human learning behavior where an agent attempts to solve a given control problem by interacting with the environment and developing an optimal policy.  Unlike the methods implemented in previous studies for solving the eco-driving problem, RL does not require prior knowledge of the environment to be learned and processed.  This paper develops a Deep Reinforcement Learning (DRL) agent for solving the Eco-Approach and Departure problem in the vicinity of signalized intersections for minimization of fuel consumption.  The DRL algorithm utilizes a Deep Neural Network for the RL.  Novel strategies such as varying actions, prioritized experience replay, target network, and double learning were implemented to overcome the expected instabilities during the training process.  The results revealed the significance of the DRL algorithm in reducing fuel consumption.  Interestingly, the DRL algorithm was able to successfully learn the environment and guide vehicles through the intersection without red light running violation.  On average, the DRL provided fuel savings of about 13.02 % with no red light running violations.

Narayana Raju, Sardar Vallabhbhai National Institute of Technology
Shriniwas Arkatkar, Sardar Vallabhbhai National Institute of Technology
Gaurang Joshi, Sardar Vallabhbhai National Institute of Technology

The article describes modeling vehicular movements using supervised machine learning algorithms with extended vehicular trajectory data prevailing heterogeneous non-lane-based traffic conditions. For this a vehicle must be tracked over a long road stretches, possibly 400 to 800m, example NGSIM data in the USA. Whereas, under heterogeneous traffic conditions prevailing in emerging countries like India, developing these kinds of vehicle trajectory data is a challenging task. On these lines, previous studies have reported a trajectory datasets usage in the range of 120m to 230m in these traffic conditions. In addressing the research gap, an algorithm was conceptualized to map the continuous vehicular movement. Based on this, trajectory data on the midblock road section is developed for a trap length around 540m at two different traffic flow conditions. Further considering the potentiality of the artificial intelligence in computational sciences, in the present work supervised machine learning algorithms, which is a subset of artificial intelligence is employed in modeling the vehicular positions. A set of parameters were identified for modeling the longitudinal and lateral instincts with surrounding vehicles. Based on the set of parameters, the algorithms compatibility in mimicking the vehicular positions are evaluated. It was identified that, considered supervised machine learning algorithms will able to model the vehicles poisitions with an accuracy in the range of 20 to 60 Mean absolute percentage error. In that k-NN algorithm found to be marginally edging past all algorithms and acting as a budding candidate for modeling vehicular positions.

Tianyang Han, University of Tokyo
Takashi Oguchi, University of Tokyo
Shiyi Liu, University of Virginia

With the development of intelligent transportation system (ITS), quantities of traffic data are collected every few seconds from large-scale urban areas by multiple sensors. These data are employed in various data-driven procedure. However, incomplete data is common and unavoidable, which has a large impact on ITS. A popular theory about recovering the incomplete traffic data is to utilize the temporal-spatial dependence. Previous studies have proposed imputation method based on calculus and statistics with matrices and high-order tensors. However, how to apply the non-Euclidean features in traffic data (e.g. road topology) is still inconclusive. In this study, we propose an imputation algorithm based on matrix decomposition method with graph regularization. The decomposition procedure can extract temporal-spatial patterns from the traffic data matrix, while graph regularization can make road topology and traffic flow periodicity understandable by imputation procedure. The numerical results show that the proposed method can be applicable for large-scale traffic speed and volume data. The comprehensive accuracy is better than both original non-negative matrix factorization (NMF) and tensor decomposition. Also, for both random and non-random missing scenarios, it shows higher robustness than NMF. This method might help in providing better data accuracy for ITS service.