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.
Incident Duration Time Prediction Using A Supervised Topic Modeling Method
Jihyun Park, Korea Expressway CorpShow Abstract
Joyoung Lee, New Jersey Institute of Technology
Branislav Dimitrijevic, New Jersey Institute of Technology
Precisely predicting the duration time of an incident is one of the most prominent components to implement proactive congestion management strategies caused by an incident. This paper presents a novel method to predict incident duration time in a timely manner by using an emerging supervised topic modeling method. Based on Natural Language Processing (NLP) techniques, this paper performs semantic text analyses with text-based incident dataset to train the model. The model is trained with actual 1,466 incident records collected by Korea Expressway Corporation from 2016 to 2019 by applying a Labeled Latent Dirichlet Allocation(L-LDA) approach. For the training, this paper divides the incident duration times into two groups: shorter than 2-hour and longer than 2-hour, based on the MUTCD incident management guideline. The model is tested with randomly selected incident records that have not been used for the training. The results demonstrate that the overall prediction accuracies are approximately 77% and 81% for the incidents shorter and longer than 2-hour, respectively.
Automated Detection and Quantification of Cracks and Spalls in Concrete Bridge Decks Using Deep Learning
Qianyun Zhang, University of PittsburghShow Abstract
Saeed Babanajad, Wiss, Janney, Elstner Associates, Inc.
Franklin Moon, Rutgers University
Amir Alavi, University of Pittsburgh
This paper presents a deep learning approach for automated detection and quantification of cracks and spalls in concrete bridge decks. The proposed concrete defect detection approach is based on the integration of convolutional neural network with a long short-term memory architecture. Thousands of manually labeled images collected from the concrete structures in Pittsburgh are used to calibrate the deep learning algorithm. The results indicate that the developed deep learning algorithm is capable of identifying the cracks and spalls on the concrete surface with acceptable accuracy. A calculation procedure is developed to quantify the density of the cracks and spalled areas that is required in the current Pennsylvania Department of Transportation (PennDOT) condition rating system for concrete bridge decks. Furthermore, a software program is designed to facilitate the implementation of the proposed method. The fast implementation of the developed deep learning framework makes it a promising tool for automated and real-time bridge and pavement inspections.
A route choice model based on cellular automaton and bounded rationality: Empirical analysis of transportation network in Sichuan-Tibet region
Junxiang XU, Southwest Jiaotong UniversityShow Abstract
Jin Zhang, Southwest Jiaotong University
Jingni Guo, Southwest Jiaotong University
In order to study the influence of travelers’ self-adaptive adjustment behavior on transportation network under the assumption of bounded rationality, a travel route choice model with individual interactive mechanism is established by using cumulative prospect theory and cellular automaton mode. In the model, travelers are divided into risk-seeking type and risk-averse-type, taking the generalized travel cost as the reference point, the choice rules of heterogeneous reference points are proposed in the study, and the evolution rules of dynamic reference points with heterogeneous characteristics are designed based on the idea of cellular genetic algorithm, so that travelers can dynamically adjust their generalized travel cost budget according to the changes of decision environment. Finally, the improved method of successive average algorithm is designed to solve the network assignment based on bounded rationality. We take Sichuan-Tibet region as our study object, mainly focus on the multi-modal transportation network in that region, and apply model and algorithm proposed in this study into travel route choice analysis and traffic assignment. It is found that the cellular automaton model can simulate the dynamic change process of the travel reference point very well, and the route choice model established by cumulative prospect theory and cellular automaton has practical significance. Such a conclusion is obtained from the empirical analysis. We find that the types of risk attitude of travelers under the bounded rational behavior will greatly affect the results of traffic assignment.
Lightweight Convolutional Neural Networks for Crowd Density Estimation
Shuo Wang (firstname.lastname@example.org), University of WashingtonShow Abstract
Ziyuan Pu, University of Washington
Qianxia Cao, Changsha University of Science and Technology
Qianmu Li, Nanjing University
Yinhai Wang, University of Washington
Crowd stepemdes and incidents is one of cardinal threat to public security that caused countless deaths during the past decades. To avoid crowd stepemdes, real-time crowd density estimation can help with monitoring crowd movement patterns, and thus supporting timely evacuation stretagy development. In previous studies, scholars and engineers developed multiple video-based crowd density estimation algorithms based on deep neural networks. The excessive computational complexity of deep learing algorithms exacerbated the algorithm’s efficiency, turning in unacceptable real-time performance. In Internet of Things era, deploy the crowd density estimation task with edge computing is an acceleration strategy to maintain real-time perfromece of the entire system. Considering the limited computational resources on the IoT device, deep learning-based crowd density estimation algorithms normally cannot be handled. To fulfill the deployment on IoT device, the algorithms need to be optimized with smaller model size. Among all deep learning methods, CNN models get less model complexity which is more suitable for developing lightweight crowd density estimaton algorithm. CNN model optimization can further improve its inference speed. Therefore, this paper proposes a lightweight CNN based crowd density estimation model by combining the modified model based on MobileNetv2 and the dilated convolution. Public crowd image data sets are used to conduct experiments for evaluating the performance of the proposed algorithm in terms of accuracy and inference speed. The results show that our model achieves much better inference speed accompanying with a slight increase in accuracy.
How Does Machine Learning Compare To Conventional Econometrics For Transport Data Sets? A Test Of ML Vs MLE
Weijia (Vivian) Li, Chang'an UniversityShow Abstract
Kara M. Kockelman (email@example.com), University of Texas, Austin
Machine learning (ML) is being used regularly in transportation and other applications to predict response variables as a function of many inputs. This paper compares traditional econometric methods that have better explanations of data analysis to ML methods, focusing on predicting, understanding and unpacking ML methods which have higher prediction accuracies of four key transport planning variables: household vehicle-miles traveled (continuous variable), household vehicle ownership (count variable), mode choice (categorical variable), and land use change (categorical variable with strong spatial interactions). Here, the results of decision trees (DTs), random forests (RFs), Adaboost, gradient boosting decision trees (GBDTs), XGBoost, lightGBM (gradient boosting methods), catboost, neural networks (NNs), support vector methods (SVMs) and Naïve Bayesian networks (BN) are compared to methods of ordinary least squares (OLS), multinomial logit (MNL), negative binomial and spatial auto-regressive (SAR) MNL methods using the U.S.’s 2017 National Household Travel Survey (NHTS 2017) and land use data sets from the Dallas-Ft. Worth region of Texas. Results suggest traditional econometric methods work pretty well on the more continuous responses (VMT and vehicle ownership), but the RF, GBDT and XGBoost methods delivered the best results, though the RF model required 30 to almost 60 times more computing time than XGBoost and GBDT methods. The RF, GBDT, XGBoost, lightGBM and catboost offer better results than other methods for the two “classification” cases (mode choice and land use change), with lightGBM being the most time-efficient. Importantly, ML methods captured the plateauing effect modelers may expect when extrapolating covariate effects.
Predicting Cycle-Level Traffic Movements at Signalized Intersections Using Machine Learning Models
Nada Mahmoud (firstname.lastname@example.org), University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
Qing Cai, University of Central Florida
Jinghui Yuan, University of Central Florida
Predicting accurate traffic parameters is fundamental and cost-effective in providing traffic applications with required information. Many studies adopted various parametric and machine learning techniques to predict traffic parameters such as travel time, speed, and traffic volume. Machine learning techniques have achieved promising results in predicting traffic volume. However, the utilized data was mostly aggregated in 5, 10, or 15 minutes. This study attempts to bridge the research gap by predicting signal cycle-level through and left-turn movements in real-time at signalized intersections. The utilized data was limited to the upstream and downstream intersections at the corridor level. Aiming to achieve this objective, eXtreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models were developed using datsets that contain variables from different number of utilized cycles (4, 6, and 8 cycles). The three models were evaluated by calculating Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). The results showed that the performance measures for the three models were close. Meanwhile. the GRU model using variables from six previous cycles outperformed the others. This modelling approach was followed to predict traffic movements for different time horizons (five cycles ahead). The performance measures values were close for the five predicted cycles. It is expected that the model could help in obtaining accurate traffic movement at intersections, which could be used for adjusting adaptive signal timing and improve signal and intersections’ efficiency.
A deep learning approach for evaluating the impacts of traffic states on crash risk in the vicinities of weaving segments
jingya zhao (email@example.com), Southeast UniversityShow Abstract
Pan Liu, Southeast University
Chengcheng Xu, Southeast University
The primary objective of this study is to divide traffic flow of type A weaving segments into different traffic states using deep learning theory, and to evaluate the impacts of different traffic states on crash risk in the vicinities of weaving segments. Using traffic flow data and crash data collected from 21 weaving segments on 12 freeways in the state of California, United States, a deep clustering with convolutional autoencoders (DCEC) is developed to classify traffic flow into ten traffic states. Comparison results suggest that the proposed deep clustering with convolutional autoencoders has the advantage in extracting the spatiotemporal features of traffic flow data and provides better cluster performance. On the other hand, the characteristics of the proposed ten traffic states are described by different lanes. The visualization of features extracted from the DCEC indicates that the spatiotemporal features of the ten traffic states are different from each other. According to the logistic regression model between traffic states and crash risk, it can be concluded that the ten traffic states are significantly associated with crash risk in the vicinities of weaving segments and each traffic state can be assigned with a certain safety level. By assigning different labels to the crashes in the ten traffic states, the CNN with gated convolutional layers (G-CNN) is further developed to predict crash risk under different traffic states. The results indicate that the propose ten traffic states can better capture the spatiotemporal features of traffic flow before crash occurrence in the vicinities of weaving segments and can provide better prediction performance.
Restricted Exploration Problem of Reinforcement Learning-based Traffic Signal Control Model and Development of Transferable Policy using Graph Neural Networks
Jinwon Yoon, Korea Advanced Institute of Science and Technology (KAIST)Show Abstract
Kyuree Ahn, Korea Advanced Institute of Science and Technology (KAIST)
Jinkyoo Park, Korea Advanced Institute of Science and Technology (KAIST)
Hwasoo Yeo (firstname.lastname@example.org), Korea Advanced Institute of Science and Technology (KAIST)
Reinforcement learning (RL) has emerged as an alternative approach for optimizing the traffic signal control system. However, there is an issue concerning the restricted exploration encountered when the signal control model is trained with the traffic simulation that has a predefined travel demand scenario. With the restricted exploration, the model would obtain a partially-trained policy that is valid only for the small part of the state space and not for the unexplored (‘never-before-seen’) states. Although this issue critically affects the robustness of the signal control model, however it has not been considered in the literature. Therefore, this research aims to obtain a transferable policy as an effective method to enhance the training efficiency when the model has a partially-trained policy due to the restricted exploration. The key idea is to represent the state variable as a graph and train it using graph neural networks (GNNs). Then, the policy can infer the solution for an unexplored state by using the already-trained knowledge of the topologically equivalent state. The experiment is conducted with five test demand scenarios of different levels in order to investigate the transferability of the policy. The results show that the proposed GNN-based model transferably adapts to the changes of traffic states than the model that does not consider the graph representation.
Short-term Forecasting of on-street Parking Occupancy Using Multivariate Spatiotemporal Variable Based on Long and Short Memory
Mengqi Lv (email@example.com), Southeast UniversityShow Abstract
Yanjie Ji, Southeast University
Shuichao Zhang, Ningbo University of Technology
Crowded streets are a critical problem in large cities, because a lot of drivers tend to seek on-street parking lots. Their selection of parking lots is affected by multiple factors, such as walking distance, trip purpose and parking price. On the other hand, parking occupancy rates have temporal and spatial dependencies. Therefore, the Random Forest model is employed to rank the importance of multiple Points of Information (POI), and the ranking is then used for feature selection. Located in the road networks, on-street parking lots have a higher impact of connectivity, which is expressed by segment incidence matrix between parking lots, on parking occupancy than that of linear distance. The long short-term memory (LSTM) that can address the time dependencies is applied to forecast the parking occupancy rates based on real parking data, and the results show that the LSTM that takes POI data into account (LSTM(+POI)) can achieves a better prediction performance than other approaches. Furthermore, during the peak hours, LSTM(+POI) can deliver a better prediction performance than LSTM. When there is a higher distribution of restaurants on both sides of streets, the prediction performance of LSTM(+POI) is significantly better than that of LSTM.
A Deep Q-learning Method for Optimal Dynamic Privileged Parking Permit Policy
Yun Yuan, University of UtahShow Abstract
Xin Wang, University of Wisconsin, Madison
Xin Li, Dalian Maritime University
Xianfeng Yang (firstname.lastname@example.org), University of Utah
Smart parking techniques continue to evolve when an increasing number of cities struggle with traffic congestion and inadequate parking availability. The privileged permit service can be provided as an alternative to the conventional meter and reserved services in the off-street parking lots. The studied privileged permit is defined as a parking product with a fixed paid period and a commitment of guaranteed available spaces. In view of the unbalanced demand and the simplistic off-street parking lot management, this paper proposes a novel parking management problem for setting-up and withdrawing the temporary permit-only policy. To optimize the access rule regarding uncertainty demand on the time of day and the utilization of the parking lot, a deep Q learning method is proposed to address the uncertainty and dimensionality in the framework of deep reinforcement learning (DRL). To replicate real world demand pattern, the proposed method integrates the long-short term memory (LSTM) for short-term demand prediction, and multivariant Gaussian process (GP) for estimating correlation in the embedded simulated environment for training deep Q network. To speed up the neural network training procedure, the proposed all-on-GPU implementation framework is $40$ times faster than the traditional hybrid-GPU framework in the DRL training. A case study is performed on urban parking lots on university campus. The numerical experiments of a rule-based strategy, a tabular Q-learning method, and the proposed deep Q-learning method are conducted to justify the effectiveness of the proposed method. The proposed method outperforms the static $(s,S)$ inventory policy by 65\% and tabular Q learning with linear Q value estimation by 15\%. The sensitivity analyses show the deep Q-learning method is capable to handle capacity-reduced, demand-increased, and special-event scenarios while the comparable strategy underperforms the proposed method.
Sensor-based Transportation Mode Recognition Using Variational Autoencoder: Application of Smart Phone Data in the Context of P2V
Zubayer Islam (email@example.com), University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
We present data augmentation technique that can improve the classification of transportation modes when the training data is insufficient. The proposed method uses a variational autoencoder (VAE) based synthetic data generation algorithm for smartphone data. Often the data collected by individuals for research is limited due to practical constraints. The algorithm discussed would aid in generating similar data from a handful of collected data to give a substantial dataset for any machine learning models. We propose a VAE, the decoder of which can help generate this synthetic data. We show that the synthetic data closely follows the pattern of the real data. We also show that classification accuracy is improved with the use of this type of data. Our method would also be a useful tool to boost the samples of an underrepresented class in a dataset. The detection of activity recognition using smartphone sensors could be applied to multiple aspects of vehicle to pedestrian P2V systems and smart mobility.
Design of an Autonomous Ride-sharing Service through a Graph Embedding Integrated Dial-a-ride Problem: Application to the Last-mile Transit in Lyon City
Omar Rifki, Ecole des Mines de Saint-EtienneShow Abstract
Nicolas Chiabaut, Universite de Lyon
Jean-Pierre Nicolas, Universite de Lyon
Autonomous vehicles are anticipated to revolutionize ride-sharing services and subsequently enhance the public transportation systems through a first-last mile transit service. Within this context, a fleet of autonomous vehicles can be modeled as a Dial-a-Ride-Problem with certain features. We propose in this study a holistic solving approach for this problem, which combines the mixed-integer linear programming formulation with a novel graph dimension reduction method based on the graph embedding framework. This latter method is effective since accounting for heterogeneous travel demands of the covered territory tends to drastically increase the size of the routing graph, thus rendering exact solving computationally infeasible. An application is provided for the real transport demand of the industrial district of “Vallée de la Chimie” in Lyon city, France. Instances involving more than 50 transport requests, and over 10 vehicles could be easily solved. Results suggest that the vehicle kilometers traveled increase with an increase of the fleet size, but at a slow pace, while the vehicle occupancy rate heavily depends on the vehicles’ capacity. Reductions in terms of GHG emissions for the applied SAV service are estimated to be around 82.4% compared to the private vehicle mode.
Optimal Travel Route Recommendation for Tourist by Ant Colony Optimization Algorithm Based on Mobile Phone Signaling Data
Haodong Sun, Beijing University of TechnologyShow Abstract
Yanyan Chen, Beijing University of Technology
Jianming Ma, Texas Department of Transportation
Xiaoming Liu, Beijing University of Technology
With the rapid development of tourism around the world, it has become of vital importance to improve the services provided to tourists to ensure the convenience and satisfaction of their travel. The optimal travel route identification and recommendation is an important part of a tourist’s plan to make the tour healthier and improve the tourist’s satisfaction and well-being, as the tourist may not be familiar with the attractions of the city to be visited. In this paper, we propose a novel research framework to help tourists make an optimal route recommendation by analyzing the historical mobile signaling data. First, we collect the mobile signaling data generated by tourists, and then crawl the city attraction location data to obtain tourists’ travel sequence. Then, we propose to employ a frequent pattern mining method to mine the popular attractions and the frequent travel sequence among attractions, then identify a tourism area, which is comprised of a cluster of attractions that are more likely to be visited on a single tour. Next, to ensure the reasonability of the recommended travel route, we adopt an Ant Colony Optimization (ACO) algorithm to find the optimal travel route among the popular attractions. Finally, an empirical study is conducted to verify the feasibility and applicability of the proposed research framework and approaches using the data from Xiamen, Fujian Province. The results of this empirical study indicate that the proposed approaches have significant potential for identifying optimal travel routes from mobile signaling data.
An Interactive Covid-19 Mobility Impact and Social Distancing Analysis Platform
Lei Zhang, University of Maryland, College ParkShow Abstract
Sepehr Ghader, University of Maryland, College Park
Michael Pack, University of Maryland, College Park
Aref Darzi, University of Maryland, College Park
Qianqian Sun, University of Maryland, College Park
Aliakbar Kabiri, University of Maryland, College Park
Songhua Hu, University of Maryland
The research team has utilized privacy-protected mobile device location data, integrated with COVID-19 case data and census population data, to produce a COVID-19 impact analysis platform that can inform users about the effects of COVID-19 spread and government orders on mobility and social distancing. The platform is being updated daily, to continuously inform decision-makers about the impacts of COVID-19 on their communities using an interactive analytical tool. The research team has processed anonymized mobile device location data to identify trips and produced a set of variables including social distancing index, percentage of people staying at home, visits to work and non-work locations, out-of-town trips, and trip distance. The results are aggregated to county and state levels to protect privacy and scaled to the entire population of each county and state. The research team are making their data and findings, which are updated daily and go back to January 1, 2020, for benchmarking, available to the public in order to help public officials make informed decisions. This paper presents a summary of the platform and describes the methodology used to process data and produce the platform metrics.
Multivariate Multi-Step Train Delay Forecasting: A Hybrid LSTM-CPS Solution
Jianqing Wu, University of WollongongShow Abstract
Qiang Wu, Lanzhou University
Luping Zhou, The University of Sydney
Chen Cai, DATA61
Bo Du, University of Wollongong
Yanlong Zhai, Beijing Institute of Technology
Wei Wei, Xi'an University
Jun Shen, University of Wollongong
Qingguo Zhou, Lanzhou University
In metropolitan cities, train (e.g., subway) delays are among the most complained events by the public communities. Different from existing researches, we present a hybrid deep learning solution for predicting multi-step train delays in this paper. Firstly, we apply a real entropy to measure the time series regularity, and we find an approximate 80.5% potential predictability on train delays. Our solution uses Long Short-Term Memory (LSTM) and Critical Point Search (CPS) to generate the forecasts for train delays. The LSTM tackle the tasks for long-term predictions of running time and dwell time. The CPS utilizes the predicted values with a nominal timetable to identify the future primary and secondary delays based on the delay causes, run-time delay and dwell time delay. Finally, we demonstrate the performance of the standard LSTM and its variants applied in a novel architecture. The results show that the variants can improve upon the standard LSTM significantly when compared through predicting time steps of dwell time feature. The experiments also show historical trend volatility with a lot of irregularities, which prompts further studies needed to tackle them.
Automated Deep Learning-Based Approach for Pavement Crack Detection Using an Inexpensive Global Shutter RGB-D Sensor and Raspberry Pi Computer
Pouria Asadi, University of Rhode IslandShow Abstract
Hamid Mehrabi, University of Guilan
Alireza Asadi, Islamic Azad University
Melody Ahmadi, Islamic Azad University
Pavement distress assessment is a significant step for pavement management. Automated pavement crack detection is a challenging task that has been researched for decades due to the complicated pavement conditions. Current pavement condition–assessment procedures are extensively time consuming, expensive, and labor-intensive. The primary goal of this paper is to develop a cost-effective and sufficiently accurate platform using an RGB-D sensor and deep learning detection models for automated pavement crack detection on a single-board ARM-based computer. To the best of our knowledge, for the first time, a pavement crack dataset is prepared using a global shutter RGB-D sensor mounted on a car, named PAVDIS2020. The proposed dataset comprising of 2,085 pavement crack images that are captured in a wide variety of weather and illuminance conditions with 5,587 instances of pavement cracks included in these images. The dataset is annotated by experts according to Pascal VOC. State‐of‐the‐art deep learning-based object detection models is deployed to train the convolutional neural networks crack detection model on PAVDIS2020. A unified implementation of the Faster R-CNN and SSD meta-architecture-based models is implemented to evaluate the accuracy, speed, and memory usage trade-off by using various CNN backbones and varying other training parameters. The proposed pavement crack detection model was able to classify the cracks with 97.6% accuracy on PAVDIS2020 dataset. The detection model is able to locate pavement cracks patterns at the speed of 12 frame per second on a passively cooled Raspberry Pi 4 single-board computer.
Pattern Recognition from Rail Grade Crossing Fatal Crashes
Subasish Das, Texas A&M UniversityShow Abstract
Steven Lavrenz, Wayne State University
Lingtao Wu, Texas A&M University
Mohammad Jalayer, Rowan University
Xiaoqiang Kong, Texas A&M University, College Station
Rail grade crossings (RGCs) or highway-rail grade crossings are the locations where a road segment and railroad track intersect at a joint. RGCs generate a range of transport, economic, social, and environmental impacts. Safety at RGCs is one of the high-priority concerns among transportation agencies. However, little research has been performed on the patterns of key contributing factors associated with RGC crashes. The aim of this study is to identify the patterns of key contributing factors that are associated with RGC related fatal crashes. This study collected nine years (2010-2018) of fatal RGC crashes from the Fatality Analysis Reporting System (FARS) to perform the analysis. To do so, this study Taxicab Correspondence Analysis (TCA), the current analysis identifies several key clusters that trigger RGC related fatal crashes. The key clusters are extremely fatal (more than one fatality) impaired crashes during inclement weather, fatal crashes local low-speed roadways at dark with no lighting, aligned county roadways, single fatal non-impaired crashes, crashes on multilane divided roadways due to turning/stopping, and crashes on high speed state highways when negotiating a curve. The findings from the current study can be used by authorities to improve RGC safety.
Enhancing the Performance of Vehicle Passenger Detection Under Adverse Weather Conditions Using Augmented-Reality-Based Machine Learning Technology
JAEYUN LEE, GnT Solution, Inc.Show Abstract
SANGCHEOL KANG, GnT Solution, Inc.
JAEDEOK LIM, GnT Solution, Inc.
SEONG GEON KIM, GnT Solution, Inc.
Changmo Kim, University of California, Davis
As the extreme traffic congestion phenomenon common in most metropolitan areas causes unnecessary travel times, a high fuel consumption, and excessive greenhouse gas emissions, transportation agencies have implemented various strategies to mitigate traffic congestion. Managed lanes, one of the worldwide applied measures, provide benefits to road users by integrating advanced technologies such as electronic and variable tolling systems. However, those agencies already implementing or considering the implementation of the managed lane strategy seek a solution to effectively and properly charge toll rates by the vehicle occupancy and to penalize violating vehicles. Vehicle passenger detection systems (VPDSs) have been developed and evaluated worldwide, but limitations still inhibit their full implementation. In this study, the authors confirmed that the performance of the deep learning algorithm, a core VPDS technology, decreases under certain adverse weather conditions due to the lack of training data sets. The performance of the YOLOv3 model trained with a normal weather data set decreased by 6.5% when it was tested under adverse weather conditions. The authors developed augmented reality (AR) models to enhance the vehicle passenger detection accuracy (VPDA) of the VPDS by training the algorithm with AR images representing virtual adverse weather conditions. The models trained with AR image sets of various weather categories (fog, rain, and snow) attained greatly enhanced VPDAs up to 7%. The final model significantly improves the VPDAs under adverse weather conditions. The proposed models could now be implemented with the integration of a road weather information system (RWIS) under adverse weather conditions.
A Lane-Changing Prediction Method Based on Temporal Convolution Network
Yue Zhang (firstname.lastname@example.org), Tongji UniversityShow Abstract
Yajie Zou, Tongji University
Jinjun Tang, Central South University
Jian Liang, Central South University
Lane-changing is an important driving behavior and unreasonable lane changes can result in potentially dangerous traffic collisions. Advanced Driver Assistance System (ADAS) can assist drivers to change lanes safely and efficiently. To capture the stochastic time series of lane-changing behavior, this study proposes a temporal convolutional network (TCN) to predict the long-term lane-changing trajectory and behavior. In addition, the convolutional neural network (CNN) and recurrent neural network (RNN) methods are considered as the benchmark models to demonstrate the learning ability of the TCN. The lane-changing dataset was collected by the driving simulator. The prediction performance of TCN is demonstrated from three aspects: different input variables, different input dimensions and different driving scenarios. Prediction results show that the TCN can accurately predict the long-term lane-changing trajectory and driving behavior with shorter computational time compared with two benchmark models. The TCN can provide accurate lane change behavior prediction, which is one key information for the development of accurate ADAS.
Predicting Freeway Traffic Speed using Attention-Based Multi-Encoder-Decoder Neural Networks
Amr Abdelraouf (email@example.com), University of Central FloridaShow Abstract
Mohamed Abdel-Aty, University of Central Florida
Jinghui Yuan, University of Central Florida
Speed prediction is a crucial yet complicated task for intelligent transportation systems. The challenge derives from the complex spatiotemporal dependencies of traffic parameters. In the past few years, deep neural networks have achieved the best traffic speed prediction performance. However, most models depend on short-term input sequences to predict short/long-term traffic speed (e.g. predicting speed for the next hour using data from the past hour). These models fail to consider the daily and weekly periodic behavior of traffic. Another problem posed by neural networks is the lack of interpretability as they often operate as “black boxes”. In this paper, an attention-based multi-encoder-decoder (Att-MED) model is proposed to predict traffic speed. The model uses convolutional-LSTMs to capture the spatiotemporal relationship of multiple input sequences, namely short-term, daily and weekly traffic patterns. The model also employs an LSTM to model the output predictions sequentially. Furthermore, attention mechanism is used to weigh the contribution of each traffic sequence towards the output predictions. The proposed network architecture, when trained end-to-end, results in a superior prediction accuracy compared to baseline models. In addition to contributing towards performance, the attention mechanism creates weight values, which when visualized, provide insights into the decision-making process of the neural network, and consequently produce explainable outputs. Att-MED’s extracted attention weights highlight the contribution of daily and weekly periodic input towards speed prediction.
A Note on Random Forest Visualization Tools in Post-Hoc Interpretation of Nonparametric Real-Time Risk Assessment Models
Arash Khoda Bakhshi, University of WyomingShow Abstract
Mohamed Ahmed (firstname.lastname@example.org), University of Wyoming
The proper cognition of causal relations between crash contributing factors and crash probability is the essential prerequisite for Active Traffic Management (ATM) in developing appropriate intervention strategies. Numerous approaches, including traditional statistics as well as black-box techniques, have been used in Real-Time Risk Assessment (RTRA) studies to define these factors. However, there is still a gap between RTRA and its practical implications that mainly stems from the difficulties in results interpretation. This study intends to bridge this gap by introducing the concept of post-hoc interpretability and utilizing the Random Forest (RF) graphical tools for safety data visualization. A nonparametric statistical Crash Prediction Model (CPM) was calibrated on 402-miles Interstate 80 in Wyoming to define significant real-time traffic-related crash contributing factors. An RF, as the Crash Interpretation Model, was developed based on the defined significant factors to interpret the CPM in a post-hoc manner. Three RF visualization tools, including Partial Dependence Plot (PDP), Individual Conditional Expectation (ICE), and Accumulated Local Effect (ALE), were scrutinized to interpret the causal effect of these factors on the crash risk. The results revealed that these techniques have many advantages, disadvantages, and unanswered questions that must be noticed by ATM as well as future studies in the safety domain. PDPs must be accompanied by ICE to explain the heterogeneity across observations. ALE is the most reliable technique in one-dimensional plots for highly correlated multi-dimensional space of variables. However, there is a substantial distinction between ALE and PDP in two-dimensional plots that makes ALE an unreliable method.
Robust Ship Trajectory Restoraction via Denoised AIS Sensing Data
Xinqiang Chen (email@example.com), Shanghai Maritime UniversityShow Abstract
jun Ling, Shanghai Maritime University
Yongsheng Yang, Batumi Navigation Teaching University
Hailin Zheng, Shanghai Maritime University
Octavian Postolache, University of Lisbon
Ship trajectory provides crucial spatial-temporal maritime information for various traffic participants (e.g., ship crew, maritime management officials), which are employed to improve maritime traffic safety. The Automatic Identification System (AIS) data is considered as a type of typical yet important data sources, which is commonly used for predicting ship trajectory. Previous studies suggested that the AIS data may contain trivial noises due to unexpected data sampling interferences (e.g., AIS data transmission overloading). To address the issue, we proposed an ensemble ship trajectory reconstruction framework with data quality control and prediction modules. Firstly, the proposed framework cleansed the raw AIS dataset with steps of trajectory separation, data denoising and sample interval normalization. More specifically, the data quality control procedure firstly determined the potential noising AIS data samples from the raw data source, which were further corrected by the moving average algorithm. After that, we averaged the time interval for the AIS data series with cubic spline interpolation mode. Finally, we predicted short term ship trajectory with artificial neural network. We have verified performance of the proposed ship trajectory framework on two ships’ AIS data.
LSTM-based Human-Driven Vehicle Trajectory Prediction in a Connected and Autonomous Vehicle Environment
Lei Lin, University of RochesterShow Abstract
Siyuan Gong, Chang'an University
Srinivas Peeta, Georgia Institute of Technology (Georgia Tech)
Xia Wu, Chang'an University
The advent of connected and autonomous vehicles (CAVs) will change driving behavior and travel environment, and provide opportunities for safer, smoother, and smarter road transportation. During the transition from the current human-driven vehicles (HDVs) to a fully CAV traffic environment, the road traffic will consist of a “mixed” traffic flow of HDVs and CAVs. Equipped with multiple sensors and vehicle-to-vehicle communications, a CAV can track surrounding HDVs and receive trajectory data of other CAVs in communication range. These trajectory data can be leveraged with recent advances in deep learning methods to potentially predict the trajectories of a target HDV. Based on these predictions, CAVs can react to circumvent or mitigate traffic flow oscillations and accidents. This study develops attention-based Long Short-Term Memory (LSTM) models for HDV trajectory prediction in a mixed flow environment. The model and a few other LSTM variants are tested on the Next Generation SIMulation (NGSIM) US 101 dataset with different CAV market penetration rates (MPRs). Results illustrate that LSTM models that utilize historical trajectories from surrounding CAVs perform much better than those that ignore information even when the MPR is as low as 0.2. The attention-based LSTM models can provide more accurate multi-step trajectory predictions. Furthermore, we conduct grid-level average attention weight analysis and identify the CAVs with higher impact on the target HDV’s future trajectories.
Personalized Short Term Trajectory Prediction Considering Car-following Behaviors
Ruibin Zhao, Southwest Jiaotong UniversityShow Abstract
Ziyan Gao, Southwest Jiaotong University
Zhanbo Sun (firstname.lastname@example.org), Southwest Jiaotong University
To ensure road safety in a mixed traffic environment, accurately predicting the behavior of non-connected vehicles in car-following (CF) process is a work that should be paid attention to. In this paper, the following vehicles trajectory prediction method (F-LSTM) based on the fusion of long short-term memory (LSTM) neural networks and full velocity difference model (FVDM) is proposed. This method includes the following features: (i) the combination of driving mechanism in the traditional CF model and deep learning model; (ii) the consideration of CF features of the subject vehicle and the vehicle interactions in the road environment. In this method, the CF trajectories of the subject vehicle with a long time series is firstly calibrated with the parameters of the FVDM using a genetic algorithm as the CF features of the driver. Then, the LSTM is used to predict the CF trajectory of the subject vehicle according to the historical trajectories of the subject vehicle and the state of the surrounding road environment. Last, based on the preliminarily obtained trajectory prediction result and driver' CF feature, a personalized CF trajectory prediction method is designed by using three structures (embedding layer, concatenate layer and fully connected network) to correct the CF trajectory according to drivers' CF features. The proposed method is tested with the Next Generation Simulation (NGSIM) data on the I-80 highway dataset. Results from the numerical experiment show that the average prediction error of the proposed method is around 1.79 m² in the predicted time domain within 5 s , which outperforms the baseline LSTM method by 5%. Keywords: Car-following, Trajectory prediction, Long short-term memory (LSTM), Full velocity difference model (FVDM)
A Neural Network Optimal Model for Classification of Unclassified Vehicles in Weigh-in-Motion Traffic Data
Cheng Peng, Purdue UniversityShow Abstract
Yi Jiang, Purdue University
Shuo Li, Indiana Department of Transportation
Tommy Nantung, Indiana Department of Transportation
A weigh-in-motion (WIM) system has the capability to perform on-site vehicle classifications based on the FHWA schema. However, WIM datasets often contain a significant portion of vehicles that could not be classified into any of the 13 vehicle classes by WIM devices. Possible reasons that the WIM classifier failed to classify these vehicles are tailgating, lane changing, traffic congestion, and equipment malfunction. Analysis of unclassified vehicles was performed with WIM recorded data. A neural network model was established to determine the appropriate allocations of unclassified vehicles to truck classes. Since the number of unclassified vehicles is often fairly high, the allocations will help to improve the accuracy of truck traffic data and thus to improve pavement design. Video records of traffic streams on an interstate section and traffic data from a nearby WIM station were utilized to identify causes for vehicle misclassifications. The optimal model was developed through model algorithm design, data processing, model training, validation, robustness analysis, and video records verification. It was found that the optimal model was effective in allocating unclassified vehicles into appropriate vehicle classes. The optimal model was able to reclassify the unclassified vehicles that have non-zeros attributes with high accuracies. The optimal model provides a useful tool for properly allocating the unclassified vehicles to the FHWA specified vehicle classes. The developed allocations can be applied to appropriately allocate unclassified vehicles to vehicle classes for pavement design and would potentially increase benefit and reduce cost with reliable and realistic pavement designs.
Recovering the Association Between Unlinked Fare Machines and Stations Using Automated Fare Collection Data in Metro Systems
Pengfei Zhang, South China University of TechnologyShow Abstract
Zhenliang Ma, Monash University
Xiaoxiong Weng, South China University of Technology
Haris Koutsopoulos, Northeastern University
Data quality is the foundation of data-driven applications in transportation. Data problems such as missing and invalid data could sharply reduce the performance of the methods used in these applications. Although there exist plenty of studies related to data quality issues, they only focus on missing or invalid data caused by infrastructure failures (e.g. loop detector malfunction). In general, there is a lack of attention to data quality issues due to insufficient data management. In this paper, we propose a tensor decomposition based framework to tackle a specific data missing problem which occurs when the machine-station dictionary of AFC databases is incomplete. In such cases, there is lots of OD information loss as the impacted machines are not linked to any station. Hence, all associated transactions may miss the origin/destination information. The proposed framework recovers the dictionary by capturing features of the passenger flow passing through the unlinked fare machine. Evaluation results show that the proposed approach could recover the missing data with high accuracy even when a number of fare machines are not linked to a station. The framework could also support other beneficial applications.
Two Algorithms for Measuring Attitudes using Natural Language
Monique Stinson (email@example.com), Argonne National LaboratoryShow Abstract
Abolfazl Mohammadian, University of Illinois, Chicago
The last two decades have witnessed the emergence and widespread adoption of attitudinal factors in behavioral models, which form the foundation of passenger travel behavior modeling. Attitudinal factors, also called latent factors due to their unobservable nature, add value to models that are otherwise purely quantitative. Parallel to this trend, a veritable explosion in analytic methods for textual data has occurred in the computer science domain, notably in Natural Language Processing (NLP). NLP has seen widespread adoption for real-time applications including search engines and chatbots. The primary innovation of this study is to treat large-scale, passive textual data sources as a source of attitudinal data, which are obtained by examining frequency of word use and context-based measurements of word use. Relative differences between companies in these two metrics form the basis of our novel techniques to develop attitudinal measurements. In doing so, our study leverages recent advancements in NLP, text mining, and transportation behavioral modeling. As such, this work contributes a unique and powerful data collection methodology to the transportation data toolkit, although the methods are expected to be relevant for applications in other domains including marketing, psychology, and others. The validity of the methods is demonstrated through two applications relating to freight transportation behavioral models, for which attitudinal data are a major gap. Differences in means for companies with different logistics operations are examined, as well as their underlying strategies. Thus, this study also makes a significant contribution to the domain of freight transportation data collection.
TRAJGAIL: GENERATING URBAN VEHICLE TRAJECTORIES USING GENERATIVE ADVERSARIAL IMITATION LEARNING
Seongjin Choi, Korea Advanced Institute of Science and Technology (KAIST)Show Abstract
Jiwon Kim, University of Queensland
Minju Park, Hannam University
Hwasoo Yeo (firstname.lastname@example.org), Korea Advanced Institute of Science and Technology (KAIST)
Recently, there are an abundant amount of urban vehicle trajectory data that is collected in the urban road networks. Many previous researches use different algorithms, especially based on machine learning, to analyze the patterns of the urban vehicle trajectories. Unlike previous researches which used discriminative modelling approach, this research suggests a generative modelling approach to learn the underlying distributions of the urban vehicle trajectory data. A generative model for urban vehicle trajectory can produce synthetic vehicle trajectories similar to the real vehicle trajectories. This model can be used for vehicle trajectory reproduction and private data masking in trajectory privacy issues. This research proposes TrajGAIL; a generative adversarial imitation learning framework for urban vehicle trajectory generation. In TrajGAIL, the vehicle trajectory generation is formulated as an imitation learning problem in a partially observable Markov decision process. The model is trained by the generative adversarial framework which use the reward function from the adversarial discriminator. The model is tested with different datasets, and the performance of the model is evaluated in terms of dataset-level measures and trajectory-level measures. The proposed model showed exceptional performance compared to the baseline models.
Linear Regularization-based Analysis and Prediction of Human Mobility in the U.S. during the COVID-19 Pandemic
Meghna Chakraborty, Michigan State UniversityShow Abstract
Md Shakir Mahmud, Michigan State University
Timothy Gates, Michigan State University
Subhrajit Sinha, Pacific Northwest National Laboratory
Since the increasing spread of COVID-19 in the U.S., with currently the highest number of confirmed cases and deaths in the world, most states in the nation have enforced travel restrictions resulting in drastic reductions in mobility and travel. However, the overall impact and long-term implications of this crisis to mobility still remain uncertain. To this end, this study develops an analytical framework that determines the most significant factors impacting human mobility and travel in the U.S. during the pandemic. In particular, we use Least Absolute Shrinkage and Selection Operator (LASSO) to identify the significant variables influencing human mobility and utilize linear regularization algorithms, including Ridge, LASSO, and Elastic Net modeling techniques to model and predict human mobility and travel. State-level data were obtained from various open-access sources for the period from January 1, 2020 to June 13, 2020. The entire data set was divided into a training data-set and a test data-set and the variables selected by LASSO were used to train four different models by ordinary linear regression, Ridge regression, LASSO and Elastic Net regression algorithms, using the training data-set. Finally, the prediction accuracy of the developed models was examined on the test data. The results indicate that among all models, the Ridge regression provides the most superior performance with the least error, while both LASSO and Elastic Net performed better than the ordinary linear model.
A DRL-based Multiagent Cooperative Control Framework for CAV Networks: a Graphic Convolution Q Network
Jiqian Dong, Purdue UniversityShow Abstract
Sikai Chen (email@example.com), Purdue University
Paul (Young Joun) Ha, Purdue University
Runjia Du, Purdue University
Yujie Li, Purdue University
Samuel Labi, Purdue University
Connected Autonomous Vehicle (CAV) Network can be defined as a collection of CAVs operating at different locations on a multi-lane corridor, which provides a platform to facilitate the dissemination of operational information as well as control instructions. Cooperation is crucial in CAV operating systems since it can greatly enhance operation in terms of safety and mobility, and high-level cooperation between CAVs can be expected by jointly plan and control within CAV network. However, due to the highly dynamic and combinatory nature such as dynamic number of agents (CAVs) and exponentially growing joint action space in a multiagent driving task, achieving cooperative control is NP hard and cannot be governed by any simple rule-based methods. In addition, existing literature contains abundant information on autonomous driving’s sensing technology and control logic but relatively little guidance on how to fuse the information acquired from collaborative sensing and build decision processor on top of fused information. In this paper, a novel Deep Reinforcement Learning (DRL) based approach combining Graphic Convolution Neural Network (GCN) and Deep Q Network (DQN), namely Graphic Convolution Q network (GCQ) is proposed as the information fusion module and decision processor. The proposed model can aggregate the information acquired from collaborative sensing and output safe and cooperative lane changing decisions for multiple CAVs so that individual intention can be satisfied even under a highly dynamic and partially observed mixed traffic. The proposed algorithm can be deployed on centralized control infrastructures such as road-side units (RSU) or cloud platforms to improve the CAV operation.
Optimal Matching Time Interval Policy for Ride-Hailing Services Using Reinforcement Learning
Guoyang Qin (firstname.lastname@example.org), Tongji UniversityShow Abstract
Qi Luo, Cornell University
Yafeng Yin, University of Michigan, Ann Arbor
Jian Sun, Tongji University
Jieping Ye, DiDi AI Labs
Matching trip requests and available drivers efficiently is deemed a central operational problem for ride-hailing platforms. A widely-adopted matching strategy is to formulate and solve bipartite matching problems repeatedly. The efficiency of matching can be improved substantially if the matching time interval is adaptively changed. The optimal time interval is determined by the trade-off between the delay penalty cost and the improved matching reward. Since the appropriate matching time interval depends on the ride-hailing system's supply and demand states, searching for an optimal policy is complicated and compounded with the actions taken in the past. We tailor a family of reinforcement-learning-based methods to overcome the curse of dimensionality and sparse reward problems. Besides, this work provides a remedy to spatial partition balance between the state representation error and the optimality gap of asynchronous matching. We finally examine the proposed methods with real-world taxi trajectory data and garner managerial insights into the general matching time interval policies.
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.