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.
Modeling Route Choice Behavior: A Federated Learning Approach
Yonghyeon Kweon, University of VirginiaShow Abstract
Bingrong Sun, National Renewable Energy Laboratory (NREL)
B. Brian Park, University of Virginia
While big data helps improving decision makings and model developments, it often runs into privacy concerns. An example would be retrieving drivers' origins and destinations information from smart phone navigation App for developing route choice behavior model. In an attempt to conserve privacy yet to take advantage of big data, we proposed to apply for federated learning approach that has shown promising application in next word prediction in smart phone keyboard without sending text to the server. Additional benefits of using federated learning is to save on data communications, by sending model parameters instead of entire raw data, and to distribute computational burden to each smart phone compared to main server. The results from real world route choice behavior data from about 20,000 drivers over one year showed that the proposed federated learning approach outperforms traditional global model and yet assures privacy.
Prediction of Lane Change Maneuvers using the SHRP2 Naturalistic Driving Study Data: A Machine Learning Approach
Anik Das, University of WyomingShow Abstract
Mohamed Ahmed, University of Wyoming
Accurate lane change prediction information in real-time is essential to safely operate Autonomous Vehicles (AVs) on the roadways; especially at the early stage of AVs deployment, where there will be an interaction between AVs and human-driven vehicles. This study proposed a reliable lane change prediction model considering features from vehicle kinematics, machine vision, driver, and roadway geometric characteristics using the trajectory-level SHRP2 Naturalistic Driving Study and Roadway Information Database. Several Machine Learning algorithms were trained, validated, tested, and comparatively analyzed including, Classification And Regression Trees (CART), Random Forest (RF), eXtrem Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), Support Vector Machine (SVM), K Nearest Neighbor (KNN), and Naïve Bayes (NB) based on six different sets of features. In each feature set, relevant features were extracted through a wrapper based algorithm named Boruta. The results showed that XGBoost model outperformed all other models in relation to its highest overall prediction accuracy (97%) and F1-score (95.5%) considering all features. However, the highest overall prediction accuracy of 97.3% and F1-score of 95.9% were observed in the XGBoost model based on vehicle kinematics features indicating that the developed model could also provide a promising prediction in absence of other data. Moreover, it was found that XGBoost was the only model that achieved a reliable and balanced prediction performance across all six feature sets. The proposed prediction model could help in trajectory planning for AVs and could be used to develop more reliable Advanced Driver Assistance Systems (ADAS) in a Cooperative Connected and Automated Vehicles environment.
Deep Learning to Detect Road Distress from Unmanned Aerial System Imagery
Long Truong, California State Polytechnic University, PomonaShow Abstract
Omar Mora, California State Polytechnic University, Pomona
Wen Cheng, California State Polytechnic University, Pomona
Hairui Tang, California State Polytechnic University, Pomona
Mankirat Singh, California State Polytechnic University, Pomona
Surface distress is an indication of poor or unfavorable pavement performance or signs of impending failure that can be classified into a fracture, distortion, or disintegration. To mitigate failing roadways, effective methods to detect road distress are needed. Recent studies associated with the detection of road distress are encouraging using object detection algorithms. Although current methodologies are favorable, they seem to be inefficient, time consuming and costly. For these reasons, we present a methodology that is based on the mask regions with convolutional neural networks (R-CNN) model coupled with the new object detection framework Detectron2 to train our model that utilizes roadway imagery acquired from an unmanned aerial system (UAS). In order for a comprehensive understanding of the performance of the proposed model, different settings are tested in the study. First, the deep learning models are trained based on both high- and low-resolution datasets. Second, three different backbone models are explored. Finally, a set of threshold values are employed. The results suggest that the proposed methodology and UAS imagery can rapidly and easily be used as efficient tools to detect road distress with an average precision (AP) score up to 95%.
Automatic Vehicle Counting and Tracking in Aerial Video Feeds Using Cascade R-CNN and Feature Pyramid Networks
Yomna Youssef, Zewail City of Science and TechnologyShow Abstract
Mohamed Elshenawy, Zewail City of Science and Technology
Unmanned Aerial Vehicles (UAV), or drones, are poised to solve many problems associated with data collection in complex urban environments. Drones are easy to deploy, have a greater ability to move and explore the environment and are relatively cheaper than other data collection methods. The paper presents an automated method for vehicle counting and tracking using aerial images and video streams. The method combines feature pyramid networks and a Cascade Region-Based Convolutional Neural Networks (Cascade R-CNN) architecture to enable accurate detection and classification of vehicles. The paper discusses the implementation and evaluation of the method and highlights its advantages compered to some state-of-the-art techniques.
A system of vision sensor based deep neural networks for complex driving scene analysis in support of crash risk assessment and prevention
Muhammad Monjurul Karim, Stony Brook UniversityShow Abstract
Yu Li, Stony Brook University
Ruwen Qin, Stony Brook University
Zhaozheng Yin, Stony Brook University
To assist human drivers and autonomous vehicles in assessing crash risks, driving scene analysis using dash cameras on vehicles and deep learning algorithms is of paramount importance. Although these technologies are increasingly available, driving scene analysis for this purpose still remains a challenge. This is mainly due to the lack of annotated large image datasets for analyzing crash risk indicators and crash likelihood, and the lack of an effective method to extract lots of required information from complex driving scenes. To fill the gap, this paper develops a scene analysis system. The Multi-Net of the system includes two multi-task neural networks that perform scene classification to provide four labels for each scene. The DeepLab v3 and YOLO v3 are combined by the system to detect and locate risky pedestrians and the nearest vehicles. All identified information can provide the situational awareness to autonomous vehicles or human drivers for identifying crash risks from the surrounding traffic. To address the scarcity of annotated image datasets for studying traffic crashes, two completely new datasets have been developed by this paper and made available to the public, which were proved to be effective in training the proposed deep neural networks. The paper further evaluates the performance of the Multi-Net and the efficiency of the developed system. Comprehensive scene analysis is further illustrated with representative examples. Results demonstrate the effectiveness of the developed system and datasets for driving scene analysis, and their supportiveness for crash risk assessment and crash prevention.
A Deep Learning Model for Off-ramp Hourly Flow Estimation
Amir Nohekhan, University of Maryland, College ParkShow Abstract
Sara Zahedian, University of Maryland, College Park
Ali Haghani, University of Maryland, College Park
This paper aims to study the traffic volume of freeway off-ramps. Freeways are the main corridors in a transportation network serving a large portion of the traffic volume. Generally, this traffic passes into the lower level roads through off-ramps. Therefore, the traffic condition of the off-ramps is an essential factor affecting the transportation network operation. Among various traffic measures, traffic volume is the most challenging one as the continuous collection of volume data is impractical. Therefore, this study aims to estimate the ramps' hourly traffic volume using a deep learning method and explore the impacts of the input feature-space and ground-truth data collection strategies on the models' performance. The focus in this study is on traffic volume estimation of off-ramps since on-ramps and off-ramps function distinctly and should be analyzed separately. The primary data sources used in this study are volume counts, probe speeds, and infrastructure characteristics of the road segments. Through training and testing several neural network models, it became evident that the incorporation of traffic flow characteristics and infrastructure attributes of the lower-level road connected to the freeway significantly improves the off-ramp's traffic volume estimation accuracy. Further, analysis illustrated that the model could capture the temporal relationships between traffic volumes of a segment at different times; however, it was not able to establish the spatial relations between the traffic volumes of different ramps across the network.
WeatherNet: Development of a Novel Convolutional Neural Network Architecture for Trajectory-Level Weather Detection Using SHRP2 Naturalistic Driving Data
MD Nasim Khan, University of WyomingShow Abstract
Mohamed Ahmed, University of Wyoming
Driver performances could be significantly impaired in adverse weather due to poor visibility and slippery roadways. Therefore, providing drivers with accurate weather information in real-time is vital for safe driving. The state-of-practice of collecting roadway weather information is based on weather stations, which are expensive and cannot provide trajectory-level weather information. Therefore, the primary objective of this study was to develop an affordable detection system capable of providing trajectory-level weather information at road surface level in real-time. This study utilized the SHRP2 Naturalistic Driving Study (NDS) video data combined with a promising machine learning technique, called Convolutional Neural Network (CNN), to develop a weather detection model with seven weather categories: clear, light rain, heavy rain, light snow, heavy snow, distant fog, and near fog. A novel CNN architecture, named WeatherNet, were carefully crafted to achieve the weather detection task. The evaluation results based on a test dataset revealed that the WeatherNet can provide excellent performance in detecting weather conditions with an overall accuracy of 93%. The performance of the WeatherNet was also compared with five pre-trained CNN models, including AlexNet, ResNet18, GoogLeNet, ShuffleNet, and SqueezeNet, which showed that the WeatherNet can provide nearly identical performance with a significant reduction in training time. The proposed weather detection model is cost-efficient and requires less computational power and therefore, can be made widely available mainly due to the recent thriving of smartphone cameras and can be used to expand and update the current Weather-based Variable Speed Limit (VSL) systems.
Advancing Association Rule Base on Gini Impurity Statistic for P redicting Transportation Mode Choice
Jiajia Zhang, Dalian Maritime UniversityShow Abstract
Tao Feng, Eindhoven University
Zhengkui Lin (firstname.lastname@example.org), Dalian Maritime University
Harry Timmermans, Technische Universiteit, Eindhoven
Recently, machine learning approaches have been applied to predict transportation mode choice as an alternative to the more commonly used discrete choice models. General class association rules (CARs) have been introduced as a promising machine learning method, but the interpretability of the prediction results in terms of the underlying behavioral decision-making process has remained a concern. In an attempt to improve CARs, this study proposes a more advanced association rule model ( named CARGIGI ) with stronger interpretability . Based on the original CARIG approach that uses information gain ( IG ) statistic for improving the predictive accuracy , in this model, the Gini impurity (GI) statistic is used to generate new rules for improving predictive accuracy and calculate the relative importance of the variables, that of the variable levels and the weight of rules in transportation mode decision process. The weight of rules is introduced as a new pruning indicator to improve the predictive accuracy, while the relative importance of the level of a variable is used to enhance the behavioral interpretability of the results. The suggested approach is applied to the 2015 Dutch National Travel Survey. Results indicate that travel distance, OV card usage frequency, travel time, and travel purpose are the most important variables, while travel party and gender are the least important variables for predicting transportation mode choice. In addition, a 10-fold cross validation test is conducted to validate the advanced model. The results show that the newly proposed model outperform both the selected machine learning algorithms and the MNL model.
Evaluating Pedestrian Time-to-death in Fatal Pedestrian Crashes: Application of Explainable Machine Learning Using SHAP Technique
Iman Mahdinia, University of Tennessee, KnoxvilleShow Abstract
Amin Mohammadnazar, University of Tennessee, Knoxville
Asad J. Khattak, University of Tennessee
Over the last recent years, the fatality of vulnerable road users has been rising significantly. Pedestrians have been recognized as the most at-risk road users because of their low level of protection. According to the National Highway Traffic Safety Administration (NHTSA), there was a 3.4% increase in pedestrian fatalities between 2017-2018. This statistic raises concern about the safety of pedestrians. A better understanding of fatal pedestrian crashes will help safety practitioners to reduce the fatalities. While pedestrian-involved fatal crashes are often assumed to be similar in studies, instant death in a fatal crash is substantially more severe than death caused by a crash several days after the crash. Therefore, this study extracts valuable information from fatal pedestrian crashes by analyzing pedestrian time-to-death ranging from instant death to death within 30 days using the Fatality Analysis Reporting System (FARS) dataset between 2015 and 2018. This study uses emergency medical service response time as the key post-crash measure, while controlling for multiple pedestrian, driver, roadway, and environmental characteristics. To extract associations of several variables with pedestrian time-to-death, an explainable XGBoost model is developed and SHAP technique is utilized to interpret the results. The results indicate that posted speed limits, lighting conditions, crash locations, crosswalk availability, land use, pedestrian age, hit and run behavior by drivers, crash time, pedestrian behavior, pedestrian intoxication, EMS response time, pedestrian gender, and driver intoxication are the key important factors that affect survival time of pedestrians. The results and the implications are discussed in detail in the paper.
Lane-Based Traffic Arrival Pattern Estimation Using License Plate Recognition Data
Chengchuan An (email@example.com), Southeast UniversityShow Abstract
Xiaoyu "Sky" Guo, Texas A&M University, College Station
Jingxin Xia, Southeast University
Zhenbo Lu, Southeast University
Understanding of traffic arrival process and its patterns is of vital importance for delay and queue analysis at intersections. Installing advance loop detectors to detect vehicle arrivals could be costly and biased. Utilization of sampled vehicle trajectory data to reconstruct the traffic arrival flow might suffer from the limited information provided by small sample sizes. The license plate recognition (LPR) data which is commonly available at intersections in cities of China is promising to overcome such limitations. This study aims to estimate lane-based traffic arrival pattern using LPR data collected at both downstream and upstream intersections. The proposed method develops a probability model with an assumption of two-stage piecewise arrival process for upstream merge movements. Given the observations of vehicle arrivals provided by matched vehicles in LPR data, the model is able to estimate the second-based mean arrival rates to each lane at the downstream intersection over the signal cycle length of upstream intersection. The proposed method is validated using actual LPR data collected at two adjacent intersections in Kunshan City, China. The results show that the proposed method can well describe the traffic arrival process either with two-stage or uniform arrival pattern in different traffic scenarios. By comparing to a benchmark method, the proposed method is more robust and reliable to reveal actual arrival patterns under different vehicle match rates in LPR data.
Travel Behavior Modeling with Images
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)
Hongzhou Lin, Massachusetts Institute of Technology (MIT)
Jinhua Zhao, Massachusetts Institute of Technology (MIT)
Joan Walker, University of California, Berkeley
Classical travel behavioral models use numeric data as inputs; however, recent deep learning advancements offer opportunities to incorporate images into modeling. This study incorporates aerial and land use images alongside numeric data into travel behavior models, focusing on effective techniques to achieve high-quality performance. We used the 2017 National Household Travel Survey (NHTS-2017) with aerial and land use images from online sources, designed residual neural networks merging images and numeric data (MG-ResNets) as inputs, and tested the model performance of image-only ResNets and MG-ResNets on predicting automobile ownership and travel mode shares for about 18,000 U.S. census tracts. We found that sequential training, early stopping, full retraining, weight decay, and other regularization techniques are critical in combining images, sociodemographics, and built environment variables for successful modeling. Urban images processed through MG-ResNets contributed most to performance when combined with numeric data. For travel mode share outcomes, image-only models can match the performance of regularized linear regressions using a full set of sociodemographic and built environment variables from NHTS-2017. To the best of our knowledge, this study is the first to discuss techniques for incorporating urban images into travel behavior models, combine images and numeric data, and demonstrate the effectiveness of these techniques by using a large-scale dataset sampled from across the U.S.
Quantized Convolutional Neural Network for Edge-based Parking Surveillance
Yifan Zhuang, University of WashingtonShow Abstract
Ziyuan Pu, University of Washington
Hao Yang, University of Washington
Yinhai Wang (email@example.com), University of Washington
The parking space insufficiency becomes more severe during the urbanization process. Being constrained by the limited urban land resources, improving the efficiency of the usage of existing parking facilities relies more on advanced parking management strategy. Video-based parking surveillance technology, being with affluent information and easy to install, is the most popular sensing method to provide real-time parking occupancy information for supporting parking management. Currently, video-based parking occupancy detection algorithms are transformed from traditional computer vision algorithms to deep learning-based algorithms. Deep learning-based algorithms gained a considerable increase in parking occupancy detection accuracy by learning fine-grained features from video images based on way more complicated algorithm structure than traditional methods. Consequently, additional complexity of deep learning-based algorithms results in excessive high computational cost. By considering the limited computational power of video sensors, most parking surveillance system deploys video processing algorithms on the server side or cloud, which raises concerns about data transfer latency and central computation pressure. Deploying algorithms on the edge side is a potential solution to solve these problems. Considering the weak computation power on the edge side, the model quantization is an efficient way to reduce the model size and boost the inference speed while keeping the accuracy. Thus, this paper proposes a quantized parking occupancy detection system for the CPU-only edge device. The experiment results show that the quantized detection algorithm keeps high accuracy and boosts the inference speed, which is much faster than other state-of-art algorithms, on the edge device significantly.
Transportation Barriers and Cancer Patients’ Decision-making: Investigating the Role of Travel in Continue or Stop Treatments
Roya Etminani-Ghasrodashti (firstname.lastname@example.org), University of Texas, ArlingtonShow Abstract
Chen Kan, University of Texas, Arlington
Ladan Mozaffarian, University of Texas, Arlington
Transportation barriers to health care facilities influence patients' health-related decision-making. However, the impact of travel on stopping a cancer treatment remains unclear in the literature. This study aims to investigate the association between cancer patients' transportation when traveling to receive radiotherapy and chemotherapy, and their decisions towards stopping or continuing treatments. In this study, a survey was designed and conducted to collect data from cancer patients with radiotherapy (n = 335) and chemotherapy (n = 347) in the USA regarding the factors in transportation that impact their decision-making. The survey contained comprehensive questions regarding personal and health-related factors while emphasizing the role of travel behavior and travel burdens on stopping or continuing radiotherapy and chemotherapy. Furthermore, machine learning models, i.e., logistic regression, random forest, artificial neural network, and support vector machine, were employed to evaluate the contribution of factors on predicting patients' decision-making. Results reveal that lack of access to transportation have a significant impact on cancer patients' decision to stop/continue treatment. Also, limited access to private vehicles can stop radiotherapy. Although our result suggests the importance of trip frequency and trip length to healthcare providers for both radiotherapy and chemotherapy, these factors have a greater contribution in following or quitting chemotherapy treatment. Understanding the travel behavior factors that make transportation a barrier for cancer patients, would help planners clarify the type of transportation interventions needed.
Extraction of Construction Quality Requirements from Textual Specifications via Natural Language Processing
JungHo Jeon, Purdue UniversityShow Abstract
Xin Xu, Purdue University
Yuxi Zhang, Purdue University
Liu Yang, Purdue University
Hubo Cai, Purdue University
Construction inspection is an essential component of the quality assurance (QA) programs of state transportation agencies (STAs), and the guidelines for this process reside in lengthy textual specifications. In the current practice, engineers and inspectors must manually go through these documents to plan, conduct, and document their inspections, which is time-consuming, very subjective, inconsistent, and error-prone. A promising alternative to this manual process is the application of natural language processing (NLP) techniques (e.g., text parsing, sentence classification, and syntactic analysis) to automatically extract construction inspection requirements from textual documents and present them as straightforward check questions. This paper introduces an NLP-based method that 1) extracts individual sentences from the construction specification, 2) preprocesses the resulting sentences, 3) applies Word2Vec and GloVe algorithms to extract vector features, 4) uses a convolutional neural network (CNN) and recurrent neural network (RNN) to classify sentences, and 5) converts the requirement sentences into check questions via syntactic analysis. The overall methodology was assessed using the INDOT specification as a test case. Our results revealed that the CNN + GloVe combination led to the highest accuracy of 91.9% and the lowest loss of 11.7%. To further validate its use across STAs nationwide, we applied it to the construction specification of the South Carolina Department of Transportation (SCDOT) as a test case, and our average accuracy was 92.6%.
Fine-Tuning Time-of-Day Signal Timing Using Signal Performance Measures
Abolfazl Karimpour (email@example.com), University of ArizonaShow Abstract
Mohammad Razaur Rahman Shaon, University of Connecticut
Yao-Jan Wu, University of Arizona
Time-of-Day (TOD) signal control is the most widely used interval-based approach that divides a day into several breakpoint intervals and schedules separate timing plans for each interval. In practice, the TOD strategy is highly responsive to travel demand and traffic pattern change. Therefore, to keep up with the change in traffic patterns during the day, TOD breakpoint intervals should be correctly determined. Also, to guarantee the mobility and performance of the arterial due to the rapid change in the travel demand, TOD timing plans need to be retimed periodically. The whole process of TOD signal timing plan development and field implementation is a costly and labor-intensive process. Therefore, to account for these difficulties, an accurate, effective, and systematic signal-performance-measure-based approach is proposed to: correctly identify the TOD breakpoint intervals, frequently fine-tune the signal timing parameters (e.g., green splits), and predict the posterior intersection mobility performance without field implementation. The proposed approach is a stepwise procedure that only incorporates signal performance measures. To test the proposed approach, multiple intersections on a major corridor in Pima County, Arizona was selected as a case study. The results of TOD breakpoint identification indicated three TOD intervals during the day: peak, off-peak, and transition time between peak and off-peak periods should be considered for this corridor. In addition, by training a Fuzzy-Logic model, it was found that by only fine-tuning the green splits on minor and major streets, the intersection delay could be improved up to 10%.
Uncertainty Quantification with Deep Learning for Spatio-Temporal Travel Demand Prediction
Qingyi Wang, Massachusetts Institute of Technology (MIT)Show Abstract
Shenhao Wang (firstname.lastname@example.org), Massachusetts Institute of Technology (MIT)
Jinhua Zhao, Massachusetts Institute of Technology (MIT)
Uncertainty prevails in transport systems. Although recent studies adopted deep learning (DL) to predict the spatio-temporal patterns in travel demand, they largely ignored the spatio-temporal structure of the demand uncertainty. This study designs a truncated mean-variance estimation (TMVE) method for the CNN-LSTM neural network, simultaneously estimating the mean and variance to quantify the spatio-temporal uncertainty in travel demand. Our experiments target the prediction of the ride hailing ridership in Chicago, and jointly estimate its mean-variance structure using past bus, transit, and ride hailing ridership as inputs. The experiments reveal a tradeoff between the prediction interval and coverage probability, which portrays an efficient frontier but cannot provide a definitive model choice. The spatio-temporal variance of the ride hailing demand presents patterns similar to the mean values: the demand uncertainty is the most significant during the rush hours and in the urban centers. Two case studies about a high-demand urban center and a low-demand urban periphery further demonstrate the importance of considering the demand uncertainty, since the prediction intervals can successfully contain the true demand while the mean estimates systematically under-predict in both cases. Overall, this study expands the DL-based travel demand analysis by quantifying variances, introduces innovative methods to describe the spatio-temporal uncertainty structure, and empirically demonstrate the spatio-temporal variance patterns for ride hailing riderships. Future studies can expand this research by comparing the TMVE method to other algorithms and use the uncertainty structure to inform the design of resilient transportation systems.
Deep Deterministic Policy Gradient based Cooperative Platoon Logitudinal Control Strategy
Yiming Yang, Chang'an UniversityShow Abstract
Wuqi Wang, Chang'an University
Haigen Min (email@example.com), Chang'an University
Siyuan Gong, Chang'an University
Xiangmo Zhao, Chang'an University
Vehicle platoon is an important part of the next generation transportation system for its potentials to enhance the traffic throughput, fuel economy, and road safety. This paper proposes a deep reinforcement learning (DRL)-based platoon longitudinal control strategy in highway environment, which mainly solved two problems, the continuous and accurate control and string stability during platoon traveling. Three key influencing factors, including spacing, velocity and acceleration, are fully considered and satisfied by the proposed strategy. First, we modeled the platoon control and illustrated the process of the reinfoecement learning. Second, we proposed a DRL process that determines the optimal strategy for platoon longitudinal control. Particularly, a multi-objective reward function is designed, which can compromise the rewards achieved of spacing error, the deviation of velocities, and the limits of the acceleration. Third, an algorithm combining actor-critic (AC) and deep Q-network (DQN) named deep deterministic policy gradient (DDPG) is adopted for solving the platoon longitudinal control problem, which can solve the control problems on continuous state space and action space effectively. Once the DDPG-based model is well trained, it can meet the platoon longitudinal control requirements fast, and has the same control accuracy as classic control methods such as distributed model predictive control (DMPC). Extensive simulations validate the effectiveness and efficiency of our proposal in terms of learning effectiveness, control accuracy and platoon stability.
Deep Learning for the Detection and Recognition of Rail Defects in Ultrasound B-scan Images
Zhengxing Chen, Southwest Jiaotong UniversityShow Abstract
Qihang Wang, Southwest Jiaotong University
Kanghua Yang, Southwest Jiaotong University
Jidong Yao, Shanghai Dongfang Maritime Engineering
Yong Liu, China Railway Chengdu Group Co
Ping Wang, Southwest Jiaotong University
Qing He, Southwest Jiaotong University
Rail defect detection is crucial to rail operations safety. Aiming at the problem of high false alarm rates and missed detection rates in rail defect detection, this paper proposes a deep learning method using B-scan image recognition of rail defects with an improved YOLO (You Only Look Once) V3 algorithm. Specifically, the developed model can automatically position a box in B-scan images and recognize EFBWs (Electric Flash Butt Welds), normal bolt holes, BHBs (Bolt Hole Breaks), and SSCs (Shells, Spalling, or Corrugation). First, we modify the network structure of the YOLO V3 model to enlarge the receptive field of the model, thus improving the detection accuracy of the model for small scale objects. Second, B-scan image data are analyzed and standardized. The K-means clustering algorithm is used to cluster the bounding boxes in the data set to obtain 12 prior boxes. Third, we adjust the initial training parameters of the improved YOLO V3 model. Finally, the experiments are performed on 453 B-scan images as the test data-set, and the results show that the B-scan image recognition model based on the improved YOLO V3 algorithm reached high performance in terms of its precision. Additionally, the detection accuracy and efficiency are improved compared with the original model.
TrafficNet: A Deep Neural Network for Traffic Monitoring Using Distributed Fiber-Optic Sensing
Chaitanya Prasad Narisetty, NEC CorporationShow Abstract
Tomoyuki Hino, NEC Corporation
Ming-Fang Huang, NEC Laboratories America Inc
Hitoshi Sakurai, NEC Corporation
Toru Ando, NEXCO Expressway Research Institute Company Limited
Shinichiro Azuma, Central Nippon Expressway Company Limited
Distributed Fiber-Optic Sensing (DFOS) for wide-area traffic monitoring is an emerging field with far-reaching applications like congestion and trajectories detection, travel time estimation, vehicle counting etc. The most captivating aspect of DFOS is that a single sensing and processing unit can monitor traffic flow in real-time for more than 80 km while utilizing existing fiber infrastructures laid alongside roadways. This work presents a novel algorithm named TrafficNet , a deep neural network for effective extraction of traffic flow patterns using DFOS systems. Proposed TrafficNet is capable of denoising DFOS data and identifying the essential components corresponding to each traversing vehicle. TrafficNet is the first of its kind neural network developed to monitor traffic by detecting vehicle trajectories and estimating various traffic flow properties using DFOS. Experimental results indicate that TrafficNet achieves 96% accuracy for estimation of average traffic speeds as compared to existing inductive loop detectors.
Driving Behavior Detection Using Semi-supervised LSTM and Smartphone Sensors
Pei Li, University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
Zubayer Islam, University of Central Florida
Driving behavior detection is an important component of proactive traffic safety management and connected vehicle systems. Most of the existing studies used traditional supervised learning to train their models with labeled data. These methods achieved promising results but were limited by the heavy dependence on the labeled data. With the development of mobile sensing technologies, massive traffic-related data can be efficiently collected by mobile devices (e.g. smartphones, tablets, etc.). Considering data labeling is an extremely time-consuming task, this paper proposed a semi-supervised LSTM to make use of the unlabeled data. Sensor data (i.e., accelerometer and gyroscope) from smartphones were collected by different drivers with a variety of phones, vehicles, and locations. After data preparation, three Long Short-term Memory (LSTM) models were trained with the proposed Semi-supervised Learning (SSL) algorithm. Experimental results indicated that the proposed semi-supervised LSTM with a small portion of the labeled data could learn from the unlabeled data and achieve outstanding results. The comparison between semi-supervised LSTM and supervised LSTM also illustrated the good performance of the proposed methods. With much fewer labeled data, semi-supervised LSTM could achieve similar, or even better results, compared with the supervised LSTM. Moreover, the proposed method outperformed other machine learning methods (e.g. Convolutional Neural Network, XGBoost, Random Forest) in terms of precision, recall, and F1-score.
Network-wide Spatiotemporal Traffic Speed Imputation Using a Generative Adversarial Network
Garyoung Lee, Seoul National UniversityShow Abstract
Eui-Jin Kim, Seoul National University
Dong-Kyu Kim (firstname.lastname@example.org), Seoul National University
Data that are complete and accurate are the most important premises of providing reliable traffic information because they are required by most statistical analyses. However, the problem of missing data is unavoidable since the data collection system is not free of errors. Recently, deep learning approaches, which are capable of capturing the inherent features and interactions in the data, have been proposed to deal with the problem of missing data. Spatio-temporal dependencies are key for the imputation of traffic data, and color-coded traffic speed images in time-space diagrams can represent them. In this paper, we propose a multi-input deep-convolutional generative adversarial imputation network (MI-DC GAIN) to impute the network-wide traffic speed on an urban expressway in the form of speed images. The proposed method uses a convolutional neural network (CNN) to deal with spatio-temporal patterns in the speed images and GAIN to focus on the data imputation. To facilitate the training DC-GAIN, speed images reconstructed by the traffic adaptive smoothing method (TASM) were used in the multi-input structure as additional information. Findings from the experiment showed that applying CNN to the structure of GAIN can enhance the model capability of learning traffic speed images, which are enhanced further by the multi-input structure with the additional reconstructed speed images. The MI-DC GAIN achieved much better performance than benchmark models in terms of accuracy and robustness to the level-of-congestion and the missing rate.
Computational Graph-based Framework for Integrating Econometric Models and Machine Learning Algorithms in Emerging Data-Driven Analytical Environments
Taehooie Kim (email@example.com), Arizona State UniversityShow Abstract
Xuesong (Simon) Zhou, Arizona State University
Ram Pendyala, Arizona State University
In an era of big data and emergence of new technologies such as app-based ride services, there are growing opportunities for better understanding human mobility patterns from newly available data sources. Statistical models have been mainly utilized to uncover and rigorously calibrate the influence of significant factors; and machine learning algorithms have been used to explore complex patterns through improved computing efficiency for large datasets. Focusing on discrete choice modeling applications, this research aims to introduce an open-source computational graph (CG)-based modeling framework for integrating the strengths of econometric models and machine learning algorithms. Particularly, multinomial logit (MNL) and nested logit (NL) models are selected to demonstrate the performance of the proposed graph-oriented functional representation. Furthermore, the calculation of the gradient in the loglikelihood function and associated Hessian matrix is systematically accomplished using automatic differentiation (AD). Using the 2017 National Household Travel Survey data, we demonstrate the estimation results of the proposed models compared to two open-source packages, namely Biogeme and Apollo. The results indicate that the CG-based choice models can produce consistent estimates of parameters and accurate calculations for the gradients of the estimated parameters with remarkable computational efficiency.
A novel spatio-temporal feature extraction method for short-term travel time predition in an urban network
Leilei Kang, Southwest Jiaotong UniversityShow Abstract
Hao Huang, Southwest Jiaotong University
Guojing Hu, Jackson State University
Weike Lu, University of Alabama
Lan Liu, Southwest Jiaotong University
In order to improve the accuracy of short-term travel time prediction in an urban road network, a hybrid model for spatio-temporal feature extraction and prediction of urban road network travel time is proposed in this research, which combines empirical dynamic modeling (EDM) and complex networks(CN) with XGBoost prediction model. Due to the highly nonlinear and dynamic nature of travel time series, it is necessary to consider time dependence and the spatial reliance of travel time series for predicting the travel time of road networks. The dynamic feature of the travel time series can be revealed by the EDM method, a nonlinear approach based on Chaos theory. Further, the spatial characteristic of urban traffic topology can be reflected from the perspective of complex networks. To fully guarantee the reasonability and validity of spatio-temporal features, which are dug by empirical dynamic modeling and complex networks(EDMCN), for urban traffic travel time prediction, and an XGBoost prediction model is established for those characteristics. Through the in-depth exploration of the travel time and topology of a particular road network in Guiyang, the EDMCN-XGBoost prediction model's performance is verified. The results show that, compared with the single XGBoost, autoregressive moving average, artificial neural network, support vector machine, and other models, the proposed EDMCN-XGBoost prediction model presents a better performance in forecasting.
Predicting Coordinated Actuated Traffic Signal Change Times using LSTM Neural Networks
Seifeldeen Eteifa, Virginia Polytechnic Institute and State University (Virginia Tech)Show Abstract
Hesham Rakha, Virginia Polytechnic Institute and State University (Virginia Tech)
Hoda Eldardiry, Virginia Polytechnic Institute and State University (Virginia Tech)
Vehicle acceleration and deceleration maneuvers at traffic signals results in significant fuel and energy consumption levels. Green light optimal speed advisory systems require reliable estimates of signal switching times to improve vehicle fuel efficiency. Obtaining these estimates is difficult for actuated signals where the length of each green indication changes to accommodate varying traffic conditions. This study details a four-step Long Short-Term Memory deep learning-based methodology that can be used to provide reasonable switching time estimates from green to red and vice versa while being robust to missing data. The four steps are data gathering, data preparation, machine learning model tuning, and model testing and evaluation. The input to the models included controller logic, signal timing parameters, time of day, traffic state from detectors, vehicle actuation data, and pedestrian actuation data. The methodology is applied and evaluated on data from an intersection in Northern Virginia. A comparative analysis is conducted between different loss functions including the mean squared error, mean absolute error, and mean relative error used in LSTM and a new loss function is proposed. The results show that while the proposed loss function outperforms conventional loss functions in terms of overall absolute error values, the choice of the loss function is dependent on the prediction horizon. In particular, the proposed loss function is outperformed by the mean relative error for very short prediction horizons and mean squared error for very long prediction horizons.
Anisotropic Kernels for Deep Convolutional Neural Network based Traffic Speed Reconstruction
Bilal Thonnam Thodi, New York University, Abu DhabiShow Abstract
Zaid Khan, New York University, Abu Dhabi
Saif Jabari (firstname.lastname@example.org), New York University, Abu Dhabi
Monica Menendez, New York University, Abu Dhabi
We propose a traffic-oriented Deep Convolutional Neural Network (Deep CNN) to reconstruct high resolution traffic speed dynamics from sparse probe vehicle trajectories. Deep CNNs developed in the traffic reconstruction literature employ kernels that are agnostic to traffic principles, namely, they are isotropic. We develop a Deep CNN with anisotropic kernels to rectify this. We explicitly capture the cause-effect relations from the empirical traffic data and in turn produce plausible and correct (in the traffic sense) speed dynamics. We test the model using vehicle trajectory data from the Next Generation Simulation (NGSIM) program to highlight the correctness of reconstruction. The results show an improvement over the corresponding naive isotropic CNN model in quantitative metrics and visual inspection of the structure of reconstructed shockwaves. This demonstrates the value added to learning models by explicitly incorporating traffic flow properties in the model's architecture.
Mode Choice Modelling with Machine Learning: A Sequential Tour-based Approach for Addressing Imbalanced Datasets
Dimitrios Pappelis (email@example.com), University College LondonShow Abstract
Emmanouil Chaniotakis, University College London
Maria Kamargianni, University College London
The continuous progress of machine learning has introduced numerous powerful classifiers that are examined as prominent alternatives to predict travellers' mode choices. However, most classifiers fail to capture the lower market share that characterizes the minority modes of transport. Although imbalanced choice datasets are common, this has been more apparent with the emergence of new modes and mobility services, which further fragment the mode choice composition. The problem is often magnified by biased sampling and measurement errors during the data collection process. The challenge of imbalanced classification in machine learning is subject of continuous multidisciplinary research, however its extensions in mode choice modelling, remain relatively unexplored. This paper provides empirical evidence of the effect that dataset imbalance might have on prediction measures and proposes a sequential tour-based framework for addressing skewed travel diary data. The framework is applied on a dataset from the city of Thessaloniki, Greece with a total of 5646 trips, using extreme gradient boosting (XGBoost). A set of performance metrics are used for the evaluation of the developed model and the output predictions are interpreted with partial dependence plots and state-of-the-art SHAP (SHapley Additive exPlanations) based on cooperative game theory. The results indicate that incorporating sequential effects can significantly improve the model’s overall performance, especially with regards to recognition rates for the minority mode, without inducing bias within the trained classifier. Key words: Mode Choice, Machine Learning, Classification, Imbalanced, Decision Trees, XGBoost, SHAP
Mining Heterogeneous Impact of Destination Attributes in Travel Demand Forecast for Different Urban Districts: A Deep Learning Approach
Shunhua Bai, University of Texas, AustinShow Abstract
Junfeng Jiao, University of Texas, Austin
Travel demand forecast plays an important role in transportation planning. Classic models often predict people’s travel behavior based on the physical built environment in a linear fashion. Due to the development of big data and machine learning techniques, many scholars have used emerging nonlinear methods to model the impact of built environments on people’s travel behavior. However, few empirical studies discussed how the impact might vary across time and space. To fill this research gap, this study used 2019 anonymous smartphone GPS data and built a long short-term memory(LSTM) recurrent neural network(RNN) to predict the daily travel demand to 6 destinations in Austin, Texas: downtown, the university, the airport, an inner-ring point-of-interest(POI) cluster, a suburban POI cluster, and an urban fringe POI cluster. By comparing the prediction results, we found that: the overall accuracy of the LSTM model was satisfactory(RMSE 339.81); the model underestimated the traffic surge for the university in the fall semester and overestimated the demand for downtown on non-working days; the precision accuracy for POI clusters was negatively related to their adjacency to downtown; different POI clusters had cases of under- or overestimation on different occasions. This study shows that the impact of destination attributes can vary across time and space due to their heterogeneous nature. Deep learning approach such as LSTM RNN in this study can implicitly model the nonlinearity and threshold variation in the time dimension and reinforce the prediction accuracy through successive learning.
Modern Public Infrastructure Development Needs and Evaluation
Austin Dikas, Prairie View A&M UniversityShow Abstract
Sarhan Musa, Prairie View A&M University
As a nation at the forefront of science and technology, the United States should make full use of state of the art tools available to make public transportation most effective. Highways are reaching capacity and have run out of room for further lane expansion. As transportation demands of America’s ever-growing cities increases, accommodations must be made to support consumer needs. In this paper, we discuss the shortcomings of American public infrastructure and recommendations on how it can be improved by making use of Artificial Intelligence (AI) and other advanced methods used in a modern country’s infrastructure. Keywords: Public infrastructure, artificial intelligence
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.