This session presents a variety of artificial intelligence and machine learning methods and tools that have been recently applied to solve problems and improve the operation and safety in a wide spectrum of transportation applications.
Multi-stage emergency decision-making method based on cumulative prospect theory and intuitionistic fuzzy number
Junxiang XU, Southwest Jiaotong UniversityShow Abstract
Jin Zhang, Southwest Jiaotong University
Jingni Guo, Southwest Jiaotong University
Abstract: Aiming at the problem where the dynamic adjustment of reference points under the impact of decision makers’ emotions may lead to different decision-making results, this research proposes a multi-stage emergency decision-making method with the emotion updating mechanism of decision makers. Firstly, the multi-stage emergency decision-making problem under emergencies is described, and the relevant decision-making information is described in the form of intuitionistic fuzzy number. Then, a method is given for setting the dynamic reference points under the impact of decision makers’ emotions, and the cumulative prospect theory is used to calculate the scenario value of each evolutionary stage of emergencies, thereby developing an emotion updating mechanism. Next, the scenario weights of each stage are calculated, and the prospect values of alternative emergency schemes at each stage are calculated according to the expected values, input costs and start-up time values of the schemes. Furthermore, by giving the method for calculating the weight of each stage, the comprehensive prospect values of the schemes are calculated to rank them. Finally, the effectiveness of the proposed method is verified by comparing it with other methods in a case analysis.
Spatiotemporal Attention-Based Graph Convolution Network for Segment-Level Traffic Prediction
Duo Li (email@example.com), University of CambridgeShow Abstract
Joan Lasenby, University of Cambridge
Traffic prediction, as a core component of intelligent transportation systems (ITS), has been investigated thoroughly in the literature. Nevertheless, timely accurate traffic prediction still remains an open challenge due to the high nonlinearities and complex patterns of traffic flows. In addition, most of the existing traffic prediction methods focus on grid-based computing problems (e.g., crowd in-out flow prediction) and point-based computing problems (e.g., traffic detector data prediction), ignoring the segment-based traffic prediction tasks. In this study, we propose an attention-based spatiotemporal graph attention network (AST-GAT) for segment-level traffic speed prediction. In particular, a multi-head graph attention block is designed to capture the spatial dependencies among road segments. Then, a component fusion block is built for speed, volume, and weather information integration. Finally, an attention-based Long short-term memory (LSTM) block is constructed for temporal dependency learning as well as segment-based speed prediction. Experiments on a real-world dataset from the Highways England demonstrate that the proposed AST-GAT model outperforms the state-of-the-art baselines, which can provide an efficient tool for segment-based traffic prediction, and therefore fill the gap between point-based and grid-based predictions.
State-wide Traffic Volume Estimation for Non-freeway Roads Using Probe-vehicle Data and Machine Learning Methods
Yi Hou, National Renewable Energy Laboratory (NREL)Show Abstract
Venu Garikapati, National Renewable Energy Laboratory (NREL)
Christopher Hoehne, National Renewable Energy Laboratory (NREL)
Kevin Kasundra, National Renewable Energy Laboratory (NREL)
Stanley Young, National Renewable Energy Laboratory (NREL)
Traffic volume data is one of the most important metrics for accurate assessment of the performance of a transportation system. High quality traffic volume data is required to effectively assess extent of delay and congestion, detect real-time perturbations to the network, and understand traffic patterns during major weather events. Traffic volume data on freeways is typically collected through continuous count stations (CCS), while there is low traffic volume observability on non-freeway roads. Addressing this issue, this study uses a state-of-the-art machine learning method, namely XGBoost, to estimate state-wide traffic volumes on non-freeway roads using probe-vehicle data. North Carolina was chosen for this case study due to its prevalence of CCSs on non-freeway roads. The results show that the proposed algorithm can estimate hourly volumes with mean absolute error (MAE) of 57 vehs/hr and R-squared of 0.87. The method also captures the abnormal traffic patterns during hurricane Florence.
Travel and Built Environment: A Deep Learning Approach
Shenhao Wang (firstname.lastname@example.org), Massachusetts Institute of Technology (MIT)Show Abstract
Rachel Luo, Massachusetts Institute of Technology (MIT)
Xiaohu Zhang, Massachusetts Institute of Technology (MIT)
Jason Lu, Massachusetts Institute of Technology (MIT)
Hongzhou Lin, Massachusetts Institute of Technology (MIT)
Joan Walker, University of California, Berkeley
Jinhua Zhao, Massachusetts Institute of Technology (MIT)
Researchers have long handcrafted built environment (BE) features to explain travel behavior (TB). However, the handcrafted features approach can be limited by measurement difficulty, information losses, and incomplete domain expertise. To address these limitations, this study introduces a deep learning (DL) approach by mapping urban images to TBs in an end-to-end manner. We obtained TBs from the 2017 National Household Travel Survey, collected aerial and land use images, and predicted automobile counts and travel mode shares by using residual neural networks (ResNets) with the images as inputs. We found that the image-based ResNets can explain a similar proportion of the variations in travel mode shares as the regressions using handcrafted BE and sociodemographic variables combined. This similarity at the census tract-level suggests that aerial and land use images contain valuable information about both TB and sociodemographics. Image-based ResNets also significantly outperformed the handcrafted BE features, confirming that the images capture built environment information not easily reduced to summary statistics, and that the long-acknowledged modest BE&TB relationship may be caused by weaknesses of the classical method. To the best of our knowledge, this study is the first to introduce image-based DL to expand the horizon of BE&TB analysis by combining the ideas of using images as BE measurements, automatically learning BE features from urban images, and end-to-end mapping of urban images to TB. Future studies can replicate the research design to validate our findings in other contexts, and focus on generating actionable insights for the practices of urban planning.
Proposed Data Model and Prototype Software for Improving the Efficiency and Efficacy of the Transportation Research Board Annual Meeting Paper Peer Review Process
David Ory, WSPShow Abstract
Sijia Wang, WSP
Gayathri Shivaraman, WSP
The Annual Meeting of the Transportation Research Board (TRB) is attended by over 13,000 participants. The core feature of the meeting are sessions devoted to the presentation of research. Research papers are submitted to TRB and assigned to TRB committees for peer review. The TRB committees are composed of approximately 25 to 35 volunteers from academia and industry. These committees review research papers and curate worthy entries into Annual Meeting sessions during the narrow time window from the paper deadline on August 1st to the posting of the preliminary Annual Meeting agenda around November 1st. For committees that receive large numbers of papers (in the 50 to 150 range), this is a difficult task. The proposition put forward in this paper is that well-organized data and proactive analysis can improve the efficiency and fairness of the paper peer review process. A proposed data model and prototype software is presented and discussed. A demonstration of the software is included.
Race, Gender, and Income Disparity in Travel Behavior Prediction with Machine Learning
Yunhan Zheng, Massachusetts Institute of Technology (MIT)Show Abstract
Shenhao Wang (email@example.com), Massachusetts Institute of Technology (MIT)
Jinhua Zhao, Massachusetts Institute of Technology (MIT)
Although researchers increasingly adopt machine learning (ML) to model travel behavior, they predominantly focus on prediction accuracy, while largely ignore the prediction disparities embedded in the ML algorithms and their adverse social impacts. Therefore, this study introduces the perspective of prediction fairness to improve the ethical use of ML in the travel behavioral modeling with the following three steps. It firstly operationalizes prediction fairness by “equality of opportunity'”, then examines prediction disparities in travel behavioral modeling with the national household travel survey, and lastly introduces an absolute correlation regularization to mitigate the prediction disparities. The results demonstrate the prevalence of prediction disparities in travel behavior modeling, which function unfavorably against the socially disadvantaged groups, such as the ethnic minority and the low-income people. High prediction accuracy is not necessarily associated with low prediction disparities, so an exclusive focus on achieving high prediction accuracy can risk damaging the already marginalized social groups. However, the prediction disparities can be mitigated by the absolute correlation regularization, thus achieving a Pareto frontier of prediction accuracy and fairness. Overall, this study introduces the important missing dimension - prediction fairness - to transport modeling through a ML perspective, and highlights the accuracy-fairness tradeoff instead of the single dimensional focus on prediction accuracy.
A Recurrent Neural Network for Estimating Speed Using Probe Vehicle Data in Urban Area
Jae Hwan Yang, The Seoul InstituteShow Abstract
Dong-Kyu Kim, Seoul National University
Seung-Young Kho (firstname.lastname@example.org), Seoul National University
Urban traffic networks comprise a combination of various links. These networks are complicated as they have numerous intersections; thus, the use of an analytical approach or parametric models for estimating the driving speed on arterial or highway roads results in low accuracy. In this study, a model is developed for estimating the link speed using the speed data collected by a probe vehicle driven across different urban traffic links, which have interrupted flows. We discover a multimodal distribution of travel speed in the data and utilize it to separate the vehicle groups. This strategy makes it possible to obtain more detailed data, which are used to determine the traffic state and increase the accuracy of the model. The developed nonlinear model, suitable for low correlations of input data, is built based on a recurrent neural network. Moreover, this study merges three machine learning techniques to apply low correlations between link properties and speed states. The developed model lowered the mean absolute error by 35.9% on average, 46.8% for the slow state, 55.7% for the state change, and 48.0% for the sudden change over 10 km/h.
Demand-driven optimization method for shared mobility services
Cyril VEVE (email@example.com), Ecole Nationale des Travaux Publics de l'EtatShow Abstract
Nicolas Chiabaut, Universite de Lyon
Shared mobility services are announced as a game-changer in transportation and a promising solution to reduce congestion and improve the performance of urban mobility. They could prefigure the arrival of autonomous vehicles. Modeling of these new services is a real challenge, especially because existing approaches are mainly an adaptation of methods devoted to classic transportation services. Consequently, this paper introduces a new data-driven optimization method fully devoted to shared mobility service. First, the proposed approach decomposes the recurrent demand based on its spatio-temporal features to overcome the drawbacks of the existing methods. Notably, it makes it possible to consider larger instances and to build robust solutions. Thus, recurrent demand patterns are identified to capture the potential demand of shared mobility services using a tailored clustering process. Second, a variant of Dial-a-Ride Problem is implemented to design robust lines to serve this demand. Such a hybrid method makes it possible to define relatively massive transport lines while maintaining spatial and temporal proximity to users’ real demand. The method is then tested with an open-source dataset released by the New York City Taxi and Limousine Commission
Crash prediction through unified analysis of driver and vehicle volatilities: Application of 1D-Convolutional Neural Network - Long Short-Term Memory
Ramin Arvin (firstname.lastname@example.org), University of Tennessee, KnoxvilleShow Abstract
Asad J. Khattak, University of Tennessee
Hairong Qi, University of Tennessee, Knoxville
Transportation safety is highly correlated with driving behavior, especially human error playing a key role in a large portion of crashes. Modern instrumentation and computational resources allow for the monitorization of driver, vehicle, and environment to extract leading indicators of crashes from multi-dimensional data streams. To quantify variations that are beyond normal in driver behavior and vehicle kinematics, the concept of volatility is applied. The study measures driver-vehicle volatilities using naturalistic driving data. By integrating and fusing multiple data streams, i.e., driver distraction, vehicle kinematics, and driving stability in real-time, this study aims to generate useful feedback to drivers and warnings to surrounding vehicles regarding hazards. The naturalistic driving data is used which contains information on more than 3500 drivers, 7589 normal driving events, and 2004 severe events, vehicle kinematics, and driver behavior. In order to capture the local dependency and volatility in time-series data 1D-Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and 1DCNN-LSTM are applied. Vehicle kinematics, driving volatility, and impaired driving (in terms of distraction) are used as the input parameters. The results reveal that the 1DCNN-LSTM model provides the best performance, with 92.36% accuracy and prediction of 71% of crashes with a precision of 93%. Additional features are extracted with the CNN layers and temporal dependency between observations is addressed, which helps the network learn driving patterns and volatile behavior. The model can be used to monitor driving behavior in real-time and provide warnings and alerts to drivers in low-level automated vehicles, reducing their crash risk.
Linkage Problem in Mathematical Optimization of Transportation Networks
Murat Bayrak, Pennsylvania State UniversityShow Abstract
S. Ilgin Guler (email@example.com), Pennsylvania State University
Methods for identifying optimal decisions for various problems in transportation networks have been extensively studied in the literature. Depending on the size of the problem, these studies often use metaheuristic methods to find solutions for various optimization problems. However, basic metaheuristic methods, such as genetic algorithms, do not perform well for the problems with linkages between decision variables. This paper investigates the linkage problem in the optimization of capacity changing network modifications. First, the linkage problem in the location optimization of dedicated bus lanes is investigated by enumerating all possible bus lane locations in a small grid network. The results suggest that the impact of implementing a bus lane at a given location depends on where existing bus lanes are located on a network. Thus, an optimization algorithm that is capable of learning linkages between decision variables is needed for the problem. Then, the optimization performance of two genetic algorithms, a Bayesian optimization algorithm, and a population-based incremental learning algorithm is compared to each other in terms of consistency and quality of the solutions, and exploration capability. Results show that algorithms that can learn linkages between decision variables perform better than the genetic algorithms.
Modeling Anticipation and Relaxation of Lane Changing Behavior Using Deep Learning
Ke quan Chen, Southeast UniversityShow Abstract
Zhibin Li, Southeast University
Pan Liu, Southeast University
Yuxuan Wang, Southeast University
Yunxue Lu, Southeast University
Recently, modeling lane changing (LC) which is the basic but complex fundamental driving behavior, has attracted significant attention. However, most of the existing LC models are homogeneous and perform not well in capturing the anticipation and relaxation phenomena of LC maneuver. To fill this gap, we adopt long short-term memory (LSTM) network by taking the LC characteristics into account. Specifically, large quantities of trajectory data are first collected by UAV in Nanjing, China. Description analysis of LC behavior is first utilized, and it is found that the factors that affecting anticipation and relaxation processes for the same lane changer are significantly different. In this context, the LSTM-An model and LSTM-Rn model are proposed to predict the behavior of anticipation and relaxation, respectively. The original training dataset is further divided into anticipation dataset which is used to train LSTM-An model and relaxation dataset which is used to train LSTM-Rn model. In addition, we apply some tests to compare our approaches with the other two baseline models using real LC dataset. The results indicate that our models achieve the best performance for the trajectory prediction in both lateral and longitude. Moreover, the simulation results show that the LSTM-An model can precisely replicate the impact of anticipation phenomenon on target lane, and the relationship between speed and spacing of lane changer during relaxation process can be regenerated by LSTM-Re model with reasonable accuracy.
Development of a Multi-Distress Detection System for Asphalt Pavements: A Transfer Learning-based Approach
Naga Siva Pavani Peraka, Indian Institute of Technology, TirupatiShow Abstract
Krishna Prapoorna Biligiri, Indian Institute of Technology, Tirupati
Satyanarayana Kalidindi, Indian Institute of Technology, Tirupati
The major objective of this research was to develop a multi-distress detection system (MDDS) that is competent in detecting various asphalt pavement functional distresses simultaneously from video images using appropriate artificial intelligence techniques. A machine learning architecture incorporated with transfer learning-based approach was utilized to quantify multiple severity-based distresses obtained from actual pavement condition images. Eighteen distress classes were defined comprising of three levels of severity pertinent to cracking, potholes, and patch deterioration. The customized MDDS algorithm was trained and tested on 1,518 images retrieved from three different datasets. During training, MDDS attained an average loss of 1.5123, while the validation mean average precision was reported to be 87.44% after 7,900 iterations. During the training process, the customized architecture transformed the training images and segmented them into two million images that potentially enhanced the probability of prediction even when the images are spatially transformed. The model detected multiple distresses in the pavement video clip at a rate of 30 frames per second, which makes it suitable for real-time distress detection. It is envisioned that the novel real-time MDDS tested on diverse datasets could be used by roadway agencies to identify and quantify severity-based distress classes during the monitoring process itself, which ultimately reduces the time between data analysis, pavement forensic evaluation, and decision making on maintenance interventions.
A CNN-Based In-Vehicle Occupant Detection and Classification Method Using SHRP 2 Cabin Images
Ioannis Papakis (firstname.lastname@example.org), Virginia Polytechnic Institute and State University (Virginia Tech)Show Abstract
Abhijit Sarkar, Virginia Polytechnic Institute and State University (Virginia Tech)
Andrei Svetovidov, Virginia Polytechnic Institute and State University (Virginia Tech)
Jeffrey Hickman, Virginia Polytechnic Institute and State University (Virginia Tech)
Amos Abbott, Virginia Polytechnic Institute and State University (Virginia Tech)
This paper describes an approach for automatic detection and localization of driver and passengers in automobiles using in-cabin images. We used a convolutional neural network (CNN) framework and conducted experiments based on the Faster R-CNN and Cascade R-CNN detectors. Training and evaluation were performed using the Second Strategic Highway Research Program (SHRP 2) naturalistic dataset. In SHRP 2, the cabin images have been blurred to maintain privacy. After detecting occupants inside the vehicle, the system classifies the presence of a driver, front-seat passenger, or back-seat passenger. For one SHRP 2 test set, the system detected occupants with an accuracy of 94.5%. Those occupants were correctly classified as front-seat passengers with an accuracy of 97.3%, as driver with 99.5% accuracy and as back-seat passengers with 94.3% accuracy. The system performed slightly better for daytime images than for nighttime images. This work is expected to facilitate research involving interactions between drivers and passengers, particularly related to driver attention and safety. The data and code will be made available in https://github.com/VTTI/occupant_detection_classification.
Mining association rules between near-crash events and geometric and trip features through a naturalistic driving dataset
Xiaoqiang "Jack" Kong (email@example.com), Texas A&M University, College StationShow Abstract
Subasish Das, Texas A&M University
Hongmin "Tracy" Zhou, Texas A&M University
Yunlong Zhang, Texas A&M University, College Station
This research aims to explore the associations between near-crash events and road geometric features/ trip features by investigating a naturalistic driving dataset. The association rule mining apriori algorithm has been applied. To provide more insights into the near-crash behavior, this research classified the near-crash events into two severity levels: trivial near-crash event (-4.5 g < deceleration rate < -7.5 g) and non-trivial near-crash event (< -7.5 g). From the perspective of descriptive statistics, the frequency of itemset generated by the apriori algorithm suggested that near-crash events are highly associated with several factors including roadway without access control, driving during non-peak hours, roadways without a shoulder or a median, roadways with the minor arterial functional class, and roadway with a speed limit between 30 to 60 mph. By comparing the frequency of occurrence of itemset on the trivial and non-trivial near-crash events, the results indicate the length of the trip has a strong indication of the type of near-crash events. The results show non-trivial near-crash events are more likely to occur if the trips are longer than 2 hours. After applying the association rule mining algorithm, more interesting patterns for two near-crash events are generated through the rules. The main findings include: 1) trivial near-crash events are more likely to occur on a roadway has a relatively lower functional class and without median and shoulder; 2) relatively higher functional roadways with relatively wide median and shoulder could also be an intriguing combination for the non-trivial near-crash events; 3) non-trivial near-crash often occurs on long trips (more than 2 hours in this research), 4) congestions on roadways with lower functional class become a dominant rule of non-trivial near-crash event.
Using an Interpretable Machine Learning Framework to Understand the Relationship of Mobility and Reliability Indices on Truck Drivers’ Route Choices
Xiaoqiang "Jack" Kong (firstname.lastname@example.org), Texas A&M University, College StationShow Abstract
Yunlong Zhang, Texas A&M University, College Station
William Eisele, Texas A&M Transportation Institute
Xiao Xiao, Texas A&M University
Although many studies have investigated the effects of providing real-time congestion information and travel time reliability information on drivers’ routing choices, minimal work has focused on truck driver decision-making – a distinct type of driver usually mistreated as the same as passenger drivers while studying routing behavior. Travel Time Index (TTI) and the Planning Time Index (PTI) are two proven indices to measure mobility and travel time reliability, respectively. This paper explores relationships between these two indices and truck drivers’ route choice through a novel interpretable machine learning framework called the Shapley Additive ExPlanation (SHAP) framework. The authors analyzed origin-destination truck trip data from the Maryland area provided by the Maryland Department of Transportation State Highway Administration (MDOT-SHA) and INRIX. The results indicate that the two indices, total trip time and their interactions, nonlinearly influence the route choices. Truck drivers are more sensitive to real-time congestion information when the difference of mobility index (TTI) on candidate routes reached a certain threshold. Moreover, when one candidate route is more reliable than another, the route preference on a less congested route is linearly increasing. An interaction study on trip time and mobility index also found that truck drivers are more sensitive to real-time congestion information when the length of the alternate routes is a large portion of the trip. The more comprehensive understanding of truck routing behavior can help transportation professionals and researchers calibrate forecasting models and inform traveler information systems targeted to truck drivers for efficient goods movement.
FROM TWITTER TO TRAFFIC PREDICTOR: NEXT-DAY MORNING TRAFFIC PREDICTION USING SOCIAL MEDIA DATA
Weiran Yao, Carnegie Mellon UniversityShow Abstract
Sean Qian (email@example.com), Carnegie Mellon University
The effectiveness of traditional traffic prediction methods, such as autoregressive or spatio-temporal models, is often extremely limited when forecasting traffic dynamics in early morning. The reason is that traffic can break down drastically during the morning commute, and the time and duration of this break-down vary substantially from day to day. Early morning traffic is crucial to inform morning-commute traffic management, but they are generally challenging to predict in advance. In this paper, we propose to mine geotagged Twitter messages as a probing method to understand the impacts of people's work and rest patterns in the evening/night of the previous day to the morning traffic of the next day. The model is tested on freeway networks in Pittsburgh as a case study. The resulting Twitter empowered predictor is surprisingly simple and powerful. We find that, in general, the earlier people rest as indicated from Tweets, the more congested roads will be in the next morning. The occurrence of big events in the evening before, represented by higher or lower tweet sentiment than normal, often results in lower travel demand in the next morning than normal days. Besides, people's tweeting activities in the early morning are positively associated with congestion in morning peak hours. We make use of such relationships to build a predictive framework which forecasts morning congestion with people's tweeting profiles extracted before 5 am, or as late as the midnight prior to the morning. The results in the Pittsburgh region support that our framework can precisely predict morning congestion, particularly for some road segments, while its prediction performance being no worse than baseline methods on other roads. Through experiments, we demonstrate our approach considerably outperforms existing methods without Twitter message features, and it can learn meaningful representation of demand from tweeting profiles that offers additional managerial insights. The proposed Twitter-empowered predictor framework can be a promising tool for real-time traffic management and potentially extended for traffic prediction at other times of day.
Intelligent Helipad Detection And(Grad-Cam) Estimation Using Satellite Imagery
David Specht, Rowan UniversityShow Abstract
Asim Waqas, Rowan University
Ghulam Rasool, Rowan University
Charles Johnson, Federal Aviation Administration (FAA)
Nidhal Bouaynaya, Rowan University
Large available databases containing the locations of helipads are known to contain more than a few errors. This is caused due to reasons including mistakes when reporting the coordinates, the removal of helipads without notifying those that maintain these databases, and newly built helipads that are not reported. Currently the most used method for verifying the coordinates in these databases is to manually go over these locations and check if the coordinates are accurate however this can be time consuming. A better auditor can be created by using machine learning and available satellite imagery. This auditor was created by training a convolutional neural network that can identify helipads at these coordinates using the available satellite imagery. The trained network was capable of correctly distinguishing between a helipad and non-helipad with an accuracy of 95\%. This algorithm will allow for better maintenance of these helipad databases, which will help pilots be able to safely land at their intended destination.
Development of Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) Applications to Classify Pavement Cracks as Top-down, Bottom-up, and Cement-Treated Reflective Cracking
Nirmal Dhakal, Louisiana State UniversityShow Abstract
Mostafa Elseifi (firstname.lastname@example.org), Louisiana State University
Zia Zihan, Iowa State University
Zhongjie Zhang, Louisiana Department of Transportation and Development
Christophe Fillastre, Louisiana Department of Transportation and Development
Jagannath Upadhyay, State University of New York (SUNY)
Field identification of top-down cracks that initiate at the pavement surface has been challenging as they may be confused with bottom-up fatigue cracking and cement-treated (CT) reflective cracking. Pavement management engineers need to differentiate between these cracks as the treatment and repair of these distresses is noticeably different. The objective of this study was to develop Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) applications to differentiate and classify top-down, bottom-up, and cement-treated reflective cracking in in-service flexible pavements using ANN and CNN models. Input variables for the ANN model were pavement age, asphalt concrete (AC) thickness, average daily traffic (ADT), type of base, crack orientation, and crack location. The ANN network architecture consisted of three layers; an input layer of six neurons, a hidden layer of ten neurons, and a target layer of three neurons. The CNN model was developed by modifying a pre-trained network, which was fitted, tested, and validated using pavement images of known crack types. After the successful validation of the models, a windows-based application was developed for the ANN model by generating a standalone application, which would allow pavement engineers to use the application to correctly classify surface cracks. An application that combines both CNN and ANN results was also created.
A Comparative Evaluation of Established and Contemporary Deep Learning Traffic Prediction Methods
Ta Jiun Ting, University of TorontoShow Abstract
Xiaoyu Wang, University of Toronto
Islam Taha, University of Toronto
Scott Sanner, University of Toronto
Baher Abdulhai, University of Toronto
Traffic prediction is an essential component in intelligent transportation systems. Various methods have been developed to solve this challenging problem over the years, including time series models, regression models, and, more recently, deep learning models. This paper provides an unbiased comparison of these methods under a variety of settings and also addresses the critical question of whether deep learning approaches can offer significant improvements over classical machine learning methods. We used a traffic simulation model of the Greater Toronto Area to generate traffic data for a stretch of highway as well as an urban region. Using these datasets, we compared the methods under five scenarios with different prediction horizons, the presence of missing data, and the presence of traffic events unseen in the training data. Our experimental results show that deep learning methods of traffic prediction, including graph convolutional neural networks, are effective for traffic prediction. Graph convolutional neural networks with shared parameters are very compact, resistant to overfitting, and performed well in all of our experiments. However, ensemble methods such as random forest regression can generate more accurate predictions at the cost of higher resource consumption during training, which may become a challenge in large transportation networks. Overall, we found that deep learning architectures should be carefully designed by restricting the input to features with known influences on the predictions, which can guide parameter learning and improve performance.
A Novel Transfer Learning Framework for Cross-Scene Pavement Distress Detection
Yishun Li, Tongji UniversityShow Abstract
Pengyu Che, Tongji University
Chenglong Liu, Tongji University
Yuchuan Du (email@example.com), Tongji University
Jian Wang, China Mobile (Shanghai) ICT Co Ltd
Deep learning method has achieved promising results in pavement distress detection. However, the effectiveness of the training model varies according to different data and scenarios. It is time-consuming and labor-intensive to recollect labeled data and re-train a new model every time the scene changes. This paper proposes a transfer learning pipeline, which enables a distress detection model to be applied cross other un-trained scenarios. The framework consists of two main components: model transfer and data transfer. The former transfers the knowledge trained in the original model to the new scene through the domain adaptation techniques, which reduces the demand for training data by at least 25%. With the same amount of training data, the method of model transfer can improve the model accuracy by about 26.55%. The latter synthesizes the annotation data for the new scene to expand the training samples, given the limited existing data. A generative adversarial neural network is employed to perform style transfer. The results show that the generated data can be used as a supplement for training and improve the accuracy of the model. The proposed pipeline can better bring out the potential of the existing pre-trained deep learning models.
Spatiotemporal Traffic Data Imputation and Pattern Discovery with Bayesian Kernelized Probabilistic Matrix Factorization
Mengying Lei, McGill UniversityShow Abstract
Aurelie Labbe, HEC Montreal
Lijun Sun, McGill University
Spatiotemporal traffic data imputation is a fundamental challenge for data-driven transportation system analysis. In this paper, we propose a new Bayesian kernelized probabilistic matrix factorization (BKMF) model for spatiotemporal data completion, which can effectively incorporate data dependencies using Gaussian process (GP) kernel constrains. Specifically, we present a fully Bayesian framework for automatically learning model parameters as well as kernel hyperparameters based on Markov chain Monte Carlo (MCMC). By imposing the inferred GP regularizations over all rows/columns of latent components, the underlying spatiotemporal correlations/consistencies can be captured explicitly. Our model is first tested on a synthetic data to show the inference effectiveness for GP hyperparameters. Then we conduct experiments on a publicly avaliable urban traffic speed data under the random missing scenario with various missing ratios. Our results show that the proposed BKMF model obtains interpretable spatiotemporal patterns and achieves higher imputation accuracy than state-of-the-art comparing methods.
A Reinforcement Learning Method based on Graph Attention Network for Multi-Depot Vehicle Routing Problem with Soft Time Window
Ke Zhang, Tsinghua UniversityShow Abstract
Meng Li (firstname.lastname@example.org), Tsinghua University
Xi Lin, Tsinghua University
FANG HE, Tsinghua University
HUIPING LI, Tsinghua University
Multi-depot vehicle routing problem with soft time windows (MD-VRPSTW) is a practical and challenging problem in urban logistics and transportation management. Better routing and scheduling decisions can result in a higher level of customer satisfaction since more customers can be served in a shorter period of time. Over the past decade, most methods for MD-VRPSTW are heuristic algorithms which require a large amount of computational efforts. With the rapid development of logistics demands, conventional methods may fail to provide high-quality solutions within an acceptable time. In this paper, we propose a novel reinforcement learning algorithm based on graph attention network (RL-GAT) to efficiently solve the problem thatbenefited from off-line training. This method regards the vehicle routing problem as a vehicle tour decoder process, and utilizes an encoder-decoder framework with attention layers to generate tours of multiple vehicles starting from different depots iteratively. Particularly, the encoder framework employs the graph attention network to mine the complex spatial-temporal correlations within time windows. Experiments show that the proposed model consistently outperforms the genetic algorithm with little computation time on three simulated networks with different scales. Finally, the robustness of the well-trained model is validated by varying the number of depots and customers.
Estimating Pedestrian Crossing Volume at Signalized Intersections
Xiaofeng Li (email@example.com), University of ArizonaShow Abstract
Peipei Xu, University of Arizona
Yao-Jan Wu, University of Arizona
The pedestrian crossing volume is one of the most important variables used to retime the signal timing for traffic delay migration and traffic safety improvement. However, most of the existing studies only focus on long-term pedestrian volume estimation for planning purposes. To bridge the research gap, this study developed a Bayesian Additive Regression Trees (BART) model to estimate the short term pedestrian crossing volume at signalized intersections. Pedestrian-related signal controller event-based data as the time-dependent variables representing the temporal trend of pedestrian crossing volume in conjunction with point-of-interset (POI) and transit trips are used as the inputs of the BART model. 70 signalized intersections are chosen from the Pima County region to calibrate and validate the developed model. With comparing with ground-truth pedestrian crossing volume data, the developed method has an R-squared of 0.83, 0.81, and 0.71 for 60min, 30min, and 15min intervals of pedestrian crossing volume. To further evaluate the performance of the developed model, the developed BART model is used in comparison to two traditional methods (stepwise linear regression and Random Forest). The comparison results show that the BART model is superior to the other two models for hourly pedestrian crossing volume estimation. Furthermore, the calibrated model is applied at 242 signalized intersections to explore the spatial and temporal patterns by estimating the hourly eastbound or westbound and northbound or southbound pedestrian crossing volume. The developed method can provide valuable information for signal retiming.
Vehicle Trajectory Planning with Hierarchical Imitation Learning in Highway Merging Scenarios
Zhen Yang, University of MichiganShow Abstract
Yiheng Feng (firstname.lastname@example.org), Purdue University
Henry Liu, University of Michigan, Ann Arbor
Trajectory planning for connected and automated vehicles (CAVs) is a complicated task because it involves multiple objectives (e.g. mobility and safety) and tuning the parameters for the objectives is time-consuming. One solution to address this issue in CAV trajectory planning is to imitate other driving behaviors (e.g., from human drivers). In this paper, a hierarchical learning from demonstration framework is proposed for the imitation, including the high-level discrete decision-making policy derivation and the low-level continuous trajectory generation. The framework is applied in a highway merging scenario. At the high level, the decision tree classifier is proposed to identify the start point of merging maneuvers. At the low level, the maximum entropy inverse reinforcement learning is adopted to learn the parameters of the trajectory optimization problem. A trajectory generation algorithm is then derived to integrate the policies at both levels. In the numerical study, the decision tree classifier is trained with a 99.0% classification accuracy. With the learned parameters from the maximum entropy inverse reinforcement learning algorithm, the generated CAV trajectories are very close to the original trajectories in the dataset, in terms of the start point of the merging maneuver, position, speed, acceleration, and heading profiles. The average Euclidean distance metric between the original trajectories and the generated trajectories is only 0.61 m.
Freeway Traffic State Estimation Using Physics-guided Machine Learning Technique
Zhao Zhang, University of UtahShow Abstract
Yun Yuan, University of Utah
Xianfeng Yang (email@example.com), University of Utah
Traffic state estimation (TSE) is a well-known concept that reconstructs the traffic state on road segments using limited observed traffic information. Recent researches have shown a successful development of classical model-driven approaches (e.g., second-order macroscopic) and data-driven approaches (machine learning - ML), but both have their limitations. Even though the model-driven approaches could depict real-world traffic dynamics, it could potentially lead to inaccurate estimation due to traffic fluctuations. In the data-driven approaches, the acquisition of a massive amount of data is required to ensure the accurate estimation, but the limited amount of the data that can obtain in practice may not be able to highlight the importance of this approach. To address these issues, our study proposes a novel framework that combines the classical traffic flow model (mentioned as physics models) with machine learning to improve the traffic state estimation accuracy. This novel framework is termed as physics-guided machine learning (PGML), which leverages the output of the traffic physical flow model along with observational features to generate predictions using a neural network framework. To illustrate the effectiveness of the PGML to cope with the problems of freeway traffic state modeling, this paper conducts empirical studies on a real-world dataset collected from a stretch of I-215 freeway in Utah. Results show that the PGML model can perform better than the previous compatible methods, including calibrated pure physical models and pure machine learning methods, especially in terms of the estimation accuracy and input robustness.
Large-scale Freeway Traffic Volume Estimation using Crowdsourced Speed Data: A Case Study in Arizona
Adrian Cottam (firstname.lastname@example.org), University of ArizonaShow Abstract
Xiaobo Ma, University of Arizona
Xiaofeng Li, University of Arizona
Yao-Jan Wu, University of Arizona
Vehicular volume is an essential measure used by transportation departments to measure the effectiveness and status of their freeway and highway networks. Volume is commonly collected using inductive loop detectors, providing point-based volume wherever they are located. Because transportation agencies accrue additional cost when they install and maintain loop detectors, they are typically located in urban areas, and do not have a high degree of spatial coverage. Other data sources, such as crowdsourced data, have a high degree of spatial coverage, but typically do not capture vehicular volume. In this study, a deep-learning-based method is used to expand the spatial coverage of vehicular volume by predicting volume using crowdsourced data. To achieve this, a stacked Long Short-Term Memory (LSTM) model is trained to predict inductive loop detector volume for the Phoenix metropolitan area using only attributes available from crowdsourced data as the input. The model is then evaluated for accuracy using K-fold cross validation methods. The trained model is then used to predict the vehicular volume along highways and freeways throughout the entire state of Arizona. The model was able to predict with a cross-validated mean average percent error of 8.45%. The proposed method has proven a cost-effective and transferable solution to estimating large-scale traffic volume for state departments of transportation.
Deep Reinforcement Learning Approach for Improving Freeway Lane Reduction Bottlenecks Throughput Via Variable Speed Limit Control Through Connected Vehicles
Reza Vatani, Old Dominion UniversityShow Abstract
Mecit Cetin, Old Dominion University
Connected vehicles (CVs) will enable various applications to improve the efficiency of traffic flow. The focus of this paper is on improving the efficiency of a freeway bottleneck through a variable speed limit (VSL) imposed on CVs. A freeway with a lane reduction, where three lanes merge into two lanes, is modeled in a microscopic simulation environment. For determining the optimal VSLs under time varying demand, a Reinforcement Learning (RL) algorithm is proposed. The RL algorithm is implemented in the simulation environment for controlling a VSL in the upstream to manipulate the inflow of vehicles to the bottleneck to minimize delays and increase the throughput. CVs are assumed to receive VSL messages through Infrastructure-to-Vehicle (I2V) communications technologies. An Asynchronous Advantage Actor-Critic (A3C) RL algorithm is implemented to determine optimal VSL policies. Through the RL control algorithm, the speed of CVs is manipulated in the upstream of the bottleneck to avoid or minimize congestion. Various market penetration rates (MPRs) for CVs are considered in the simulation. It is demonstrated that the RL algorithm can adapt to the stochastic arrivals of CVs and achieve significant improvements even at low MPRs of CVs. The paper presents numerical experiments demonstrating the effectiveness of the RL algorithm under varying MPRs of CVs.
Unsupervised Learning to Support Early Identification of Traffic Pattern Changes: A Case Study of 2018 Heavy Rain Disaster in Hiroshima
Canh Do (email@example.com), Hiroshima UniversityShow Abstract
Makoto Chikaraishi, Hiroshima University
Akimasa Fujiwara, "Hiroshima Daigaku"
Yasuhiro Kusuhashi, West Nippon Expressway Engineering Chugoku
The dynamic monitoring and analysis of expressway traffic volume play important roles in traffic management, especially in the period of disaster events. In the information technology era, the availability of traffic sensing data provides an unprecedented opportunity for early identification of traffic pattern changes. This study proposes a systematic approach for anomaly detection from acquired traffic data based on the Bayesian estimator of abrupt change, seasonal change, trend (BEAST) method , temporal clustering analysis, and changing point analysis. The goals of the proposed approach are to support early identification of abnormal changes in traffic state to facilitate urgent policy decisions soon after disaster events. To gain an insight into these change patterns, we applied a visualization method that generates a variety of interest heatmap and line charts for traffic states over time. Using traffic data collected from 58 loop-detectors installed on expressways in the Chugoku region of Japan, the proposed method is applied to traffic volumes before, during, and after the July 2018 heavy rain disaster to exemplify how the proposed approach works. The results show that this approach can accurately identify the traffic volume change and time of disaster events. Moreover, the proposed systematic approach could be further used to develop an abnormal traffic state monitoring and detection tool or a traffic pattern classifier tool towards building a better system of network traffic management for improved resilience. It is important for engineers, planners, decision-makers, and other stakeholders to have such a tool for making timely traffic regulations and policy decisions.
Avoiding Gridlock On Large Congested Networks: A Multi-agent Deep Reinforcement Learning Approach with Spillback Knowledge
Hao Zhou (firstname.lastname@example.org), Georgia Institute of Technology (Georgia Tech)Show Abstract
Jorge Laval, Georgia Institute of Technology (Georgia Tech)
It has been recently shown that deep reinforcement learning (DRL) cannot learn useful policies under congestion mainly because urban networks seem to be insensitive to signal control. As a sequel study, here we release its assumption on adaptive routing, and revisit the DRL problem of signal control in large, oversaturated networks featuring frequent spillbacks and deadlocks. First we identify the influencing factors that deteriorate the training process of DRL, including high density levels, short block lengths, and vanishing left-turn probabilities. Accordingly, we propose a multi-agent reinforcement learning (MARL) approach and extends DRL to all density levels by: i) following a decentralized, shared learning framework which requires no explicit coordination, ii) equipping agents with spillback knowledge--a tree graph parsed from the spilled-over queues and their relations, and iii) increasing the left-turn probabilities during training. We compared the policies trained in free flow and congestion, and results show that: i) they can both well generalize to full density levels: ii) although policies are different, the corresponding MFDs are the same, regardless of driver adaptation. The study provides further evidence that the congested networks are insensitive to signal control. A random policy can be competitive to the optimal polices found in free flow or congestion.
Multi-Modal Traffic Speed Monitoring: A Real-Time System Based on Passive Wi-Fi and Bluetooth Sensing Technology
Ziyuan Pu, University of WashingtonShow Abstract
Zhiyong Cui, University of Washington
Shuo Wang, University of Washington
Hao Yang, University of Washington
Yinhai Wang (email@example.com), University of Washington
Traffic speed is one of the critical indicators reflecting traffic status of roadway networks. The abnormality and sudden changes of traffic speed indicates the occurrence of traffic congestions, accidents, and events. Traffic control and management systems usually utilize the spatio-temporal variations of traffic speed to re-schedule traffic signal timing plan, broadcast traffic events and accidents, and make management decisions. To guarantee the effectiveness of traffic control and management, accurate and real-time multi-modal traffic speed monitoring is essential for the systems aiming to control and manage roadway traffic. In previous study, Wi-Fi and Bluetooth passive sensing technology was demonstrated as an effective method for obtainining traffic speed data. However, there still some issues generat impacts on speed estimatin, e.g., traffic mode uncertainty and errors caused by detection range. Thus, this study develops a multi-modal traffic speed estimation algorithm for estimating traffic speed of the road networks covered by Wi-Fi and BT passive sensors in a real-time way. To addressing aformetioned two issues, an algorithm is developed to correct the estimated traffic speed based on Received Signal Strength Indicator of Wi-Fi and BT signals, and traffic mode identification algorithm is proposed based on a designed semi-supervised Possibilistic Fuzzy C-Means clustering algorithm. The perfromenc of the proposed algorithm is evaluated by comparing with the selected baseline algorithms. The results indicate the superiority of the proposed algorithm in speed estimation accuracy. The findings of this study can help with identifying multi-modal traffic speed simultaneously, and thus supporting traffic control and management.
Freeway Traffic State Prediction using Constructed Traffic Information from Hybrid Machine Learning
Zhao Zhang, University of UtahShow Abstract
Yun Yuan, University of Utah
Xianfeng Yang (firstname.lastname@example.org), University of Utah
Traffic state estimation (TSE) and short-term traffic state prediction are critical for the successful operation of Intelligent Transportation Systems (ITS) on freeways. However, traffic state prediction heavily relies on historical traffic information collected by various traffic sensors and the accuracy of TSE is undermined by sparse traffic sensors. To overcome this limitation, this paper introduces an innovative traffic state prediction framework that: (a) adopts the constructed traffic information acquired by the hybrid machine learning (ML) approach; and (b) uses the Long Short-Term Memory Neural Network (LSTM NN) to predict traffic state on the freeway. The hybrid ML leverages the probe vehicle data and limited observations from traffic sensors to construct precise and full-field historical traffic state. Then, by using the constructed traffic states as the input, the LSTM NN model is further developed to predict the future traffic states. To illustrate the effectiveness of the proposed traffic prediction method, this paper conducts empirical studies on real-world data collected from a stretch of I-215 freeway in Salt Lake City, Utah. Results show that the proposed method could accurately predict the freeway traffic state with the constructed historical traffic information. Also, the proposed method has been proved to capture historical traffic dynamic patterns.
DISCLAIMER: All information shared in the TRB Annual Meeting Online Program is subject to change without notice. Changes, if necessary, will be updated in the Online Program and this page is the final authority on schedule information.