|This session presents successful transportation applications of various modeling, recognition, and classification tools to traffic pattern recognition, vehicle and pedestrian recognition, connected vehicles, traffic metering, car following, travel behavior, driver behavior and response pattern, data imputing, transportation network reliability, transportation cost, delivery scheduling, shared mobility and ridesourcing, bridge performance, pavement analysis, and road surface condition classification. Tools include reinforcement learning, deep Q networks, support vector machines, gradient boosting, long short-term memory, Bayesian regularization neural networks, random forest, convolutional neural networks, and others.|
Identifying Multimodal Conflicts with Machine Learning
Nancy Hui, University of TorontoShow Abstract
Matthew Roorda, University of Toronto
Eric Miller, University of Toronto
This study explores the efficacy of using machine learning techniques to automatically identify traffic conflicts. Quantitative conflict identification methods are largely designed through observation of motorized vehicles only, and can report erroneous results when applied to users of non-motorized modes. A dataset of conflict and non-conflict events is constructed through analysis of video footage of a multimodal street. For each event, conflict indicators and parameters representing user mode, speed, and acceleration are calculated. Six machine learning classifiers are trained on 80% of the dataset: three classifiers were trained using only the conflict indicators, and three classifiers were trained using the full set of explanatory variables. Five of the six classifiers are more effective in identifying conflicts than the threshold-based conflict identification technique, suggesting that the structure of machine learning classifiers presents advantages over conventional indicator thresholds in conflict identification. Furthermore, the classifiers trained on the full set of explanatory variables performed better during conflict identification than classifiers excluding mode, speed, and acceleration in their set of potential explanatory variables. This suggests that user mode, speed, and acceleration influence interaction severity.
Automatic Background Filtering Method for Roadside Lidar Data
Jianqing Wu, University of Nevada, RenoShow Abstract
Hao Xu, University of Nevada, Reno
Yuan Sun, University of Nevada, Reno
Jianying Zheng, Soochow University
Rui Yue, University of Nevada, Reno
The high-resolution micro traffic data (HRMTD) of all roadway users is important to serve the connected-vehicle system under mix traffic situation. The roadside LiDAR sensor gives a solution to provide HRMTD by generating real-time 3D point clouds from its scanned objects. The background filtering is the basic data preprocessing to obtain the HRMTD of different roadway users. An automatic background filtering procedure was implemented to identify and exclude background points from the roadside 360o LiDAR sensor data. An algorithm is developed based on the distribution of points density in the space, which can filter both static background and moving background effectively. Different density thresholds are applied in this algorithm to filter background with different distances from the roadside sensor, which is based on the knowledge from historical data analysis. The case study shows this algorithm can be used for background filtering under different situations (different road geometries, different traffic volume; day/night time; different speed limits). Vehicle and pedestrian shape can be well kept after background filtering. The low computational load guarantees this method to be applied for the real-time data process such as vehicle monitoring and pedestrian tracking.
Rebalancing Shared Mobility-on-Demand Systems: A Reinforcement Learning Approach
Jian Wen, Massachusetts Institute of Technology (MIT)Show Abstract
Jinhua Zhao, Massachusetts Institute of Technology (MIT)
Patrick Jaillet, Massachusetts Institute of Technology (MIT)
Shared mobility-on-demand systems have very promising
prospects in making urban transportation efficient and affordable. However, due
to operational challenges among others, many mobility applications still remain
niche products. This paper addresses rebalancing needs that are critical for
effective fleet management in order to offset the inevitable imbalance of
vehicle supply and travel demand. Specifically, we propose a reinforcement
learning approach which adopts a deep Q network and adaptively moves idle
vehicles to regain balance. This innovative model-free approach takes a very
different perspective from the state-of-the-art network-based methods and is
able to cope with large-scale shared systems in real time with partial or full
data availability. We apply this approach to an agent based simulator and test
it on a London case study. Results show that, the proposed method outperforms
the local anticipatory method by reducing the fleet size by 14% while inducing
little extra vehicle distance traveled. The performance is close to the optimal
solution yet the computational speed is 2.5 times faster. Collectively, the
paper concludes that the proposed rebalancing approach is effective under
various demand scenarios and will benefit both travelers and operators if
implemented in a shared mobility-on-demand system.
Winter Road Surface Condition Recognition Using a Pretrained Deep Convolutional Neural Network
Guangyuan Pan, University of WaterlooShow Abstract
Liping Fu, University of Waterloo
Ruifan Yu, University of Waterloo
Matthew Muresan, University of Waterloo
This paper investigates the application of the latest machine learning technique – deep neural networks for classifying road surface conditions (RSC) based on images from smartphones. Traditional machine learning techniques such as support vector machine (SVM) and random forests (RF) have been attempted in literature; however, their classification performance has been less than desirable due to challenges associated with image noises caused by sunlight glare and residual salts. A deep learning model based on convolutional neural network (CNN) is proposed and evaluated for its potential to address these challenges for improved classification accuracy. In the proposed approach we introduce the idea of applying an existing CNN model that has been pre-trained using millions of images with proven high recognition accuracy. The model is extended with two additional fully-connected layers of neurons for learning the specific features of the RSC images. The whole model is then trained with a low learning rate for fine-tuning by using a small set of RSC images. Results show that the proposed model has the highest classification performance in comparison to the traditional machine learning techniques. The testing accuracy with different training dataset sizes is also analyzed, showing the potential of achieving much higher accuracy with a larger training dataset.
An Implementation-Ready Approach for Multiple-Van Multicriteria Dynamic Demand Rebalancing at Bikeshare Stations
Jiangbo Yu, University of California, IrvineShow Abstract
Dingtong Yang, University of California, Irvine
Daisik Nam, University of California, Irvine
Sunghi An, University of California, Irvine
R. Jayakrishnan, University of California, Irvine
Bike-sharing programs are increasingly popular as an effective way to enhance walk, transit, ride sharing, and car sharing accessibility. One common challenge is to find an efficient bike rebalancing strategy when pick-up and drop-off demands at bike stations are not evenly distributed in space and time. The goal of the rebalancing operation is to improve service level and reduce unsatisfied demand. Most, if not all, existing methods adopt approaches with adjustments based on spatial clustering and conventional network analysis techniques with a single criterion. This paper provides a ready-to-implement alternative to resolve the rebalancing problem with high model interpretability and tractability. The core concept of the proposed algorithm evolves from an observation that the solution set can be formed as a set of pickup-dropoff stationpairs rather than individual stations. An unsupervised learning approach is used for parameter estimation and validation. The objective function assigns weights to the cost of the bike-redistribution van operation and the cost for unsatisfied demand. The algorithm contains three general steps with feedback. The first step converts the dynamic problem into a static problem using discounting method for dynamic demand; the second step assigns bike station pairs and finds the routing of each van stochastically; the third step converts the static problem back to dynamic to determine detailed operation variables such as the exact number of bikes to serve. Heuristic search and random perturbation for bike pair sequence is utilized to avoid the solutions’ being trapped at local optima. The final validation using a training dataset and other data shows no overfitting problem for the models, and the results are consistent and efficient.
Distributed Optimization and Coordination Algorithms for Dynamic Traffic Metering in Urban Street Networks
Rasool Mohebifard, Washington State UniversityShow Abstract
Ali Hajbabaie, Washington State University
Previous research has shown that proper metering of entry traffic to urban street networks, similar to metering traffic on on-ramps in freeway facilities, reduces traffic congestion especially in oversaturated flow conditions. Building on previous research, this paper presents a real-time and scalable methodology for finding near-optimal metering rates dynamically in urban street networks. The problem is formulated into a Mixed-Integer Linear Program (MILP) based on the Cell Transmission Model (CTM). We propose a distributed optimization scheme that decomposes the network level MILP into several link-level MILPs to reduce the complexity of the problem. We convert the link-level MILPs to linear programs to reduce computational complexity further. Besides, we create distributed coordination between link-level linear programs to push solutions towards optimality. The distributed coordination and optimization solution algorithm is incorporated into a rolling horizon technique to account for stochastic demand and capacity and to further to reduce the computational complexity. We applied the proposed solution technique to a number of case studies and observed that it was scalable and real-time, and found solutions that were at most 2.2% different from the optimal solution of the problem. Like previous studies, we found significant improvements in network operations as a result of traffic metering.
A Novel Clustering Algorithm for Traffic Operational Analysis
Mohammed Almannaa, Virginia Polytechnic Institute and State UniversityShow Abstract
Mohammed Elhenawy, Virginia Polytechnic Institute and State University
Hesham Rakha, Virginia Polytechnic Institute and State University
This paper proposes a novel clustering algorithm to analyze high dimensional large datasets. The new clustering algorithm answers several questions related to traffic operations by finding the similar months-of-year, days-of-week, and time-of-day using instantaneous travel time and speed data. The proposed algorithm models the clustering problem as a matching problem between two disjoint sets of agents: centroids and data points. This new view of the clustering problem makes the proposed algorithm a multi-objective algorithm where the purity and the reciprocal of distance (similarity) in each cluster are maximized simultaneously. The proposed algorithm showed promising performance when applied to instantaneous travel time and speed data collected on I-64. When studying how traffic patterns evolve during the days of the week, the algorithm successfully grouped Wednesday, Thursday and Friday in one cluster and the rest of the weekdays (Saturday, Sunday, Monday and Tuesday) in another cluster. Additionally, when studying how traffic patterns change during seasons, the algorithm grouped the summer months in one cluster and the rest of the year in another cluster. Moreover, the proposed algorithm grouped the time-of-day into two clusters: peak hours from 7:00 a.m. to 10:00 p.m., and excluding 7:00 p.m. and off-peak hours from the rest of the hours.
Machine Learning Versus Spatial Econometric Models: Modeling the Impact of Transportation Infrastructure on Real Estate Prices
Dimitrios Efthymiou, Technical University of MunichShow Abstract
Constantinos Antoniou, Technical University of Munich
Linear regression with Ordinary Least Squares and spatial econometric models are statistical methods widely employed to measure the impact of transportation infrastructure locations on real estate prices. This paper extends the research that begun by the authors in 2013. Efthymiou and Antoniou (1, 2, 3) developed different types of OLS and spatial econometric models to measure the impact of transportation infrastructure and policies on purchase and rent prices of dwellings, using data downloaded from publicly available on-line sources. They found that spatial econometric models perform better than OLS in terms of model fit and detection of spatial autocorrelation, resulting in lower AIC and Moran’s I.
In this research the authors investigate the potential of using Machine Learning (ML) models to measure transportation cost capitalization on real estate prices, and benchmark their results versus spatial econometric models and OLS. They develop different types of random forest and gradient boosting machine models using hyperparameter optimization and stacked ensemble learning. The results show that ML models outperform traditional statistical techniques in terms of model fit (e.g. lower MSE) and successfully resolve heteroscedasticity between predicted and observed values. Moreover, it is shown that transformation of independent variables does not improve the performance of ML models. However, the output of ML models cannot be interpreted in a similar manner as of statistical models, meaning that the elasticities and percentage impact of transportation infrastructure on real estate prices cannot be measured.
Artificial Neural Network Models for Predicting Pavement Roughness of Flexible and Rigid Pavements
LEELA SAI PRAVEEN GOPISETTIShow Abstract
Mohammad Hossain, Bradley University
Suruz Miah, Bradley University
Kerrie Schattler, Bradley University
International Roughness Index (IRI) is predicted for flexible and rigid pavements using Artificial Neural Network (ANN) modeling. This study considers only climate and traffic factors as a cause of pavement roughness, means pavement distresses, and structural data are not taken into account. Climate and traffic data are used as input parameters, and the data are collected from the Long Term Pavement Performance (LTPP) database. After several trial and errors, a 7-9-9-1 ANN network is developed using two layered hyperbolic tangent sigmoid transfer function. Root Mean Square Error (RMSE) is used for comparison of LTPP measured, and ANN predicted IRI values. The prediction consistency of the ANN model is tested by expanding the pavement sites selection over wet-freeze, dry-freeze, wet no-freeze and dry no-freeze climatic zones. For both flexible and rigid pavements, the lowest 0.010 RMSE is recorded for wet no-freeze climate zone. ANN model validation is done by checking ANN generated transfer function weights and randomly generated weights. Also, chi square goodness of fit test was performed to summarize the discrepancy between observed and predicted IRI, and the results showed that for all the sections, the null hypothesis is accepted and good fits are achieved. ANN predicted roughness models would be a viable alternative for local transportation agencies those do not have access to automated vehicles to measure IRI.
Evaluation of the Gradient Boosting of Regression-Trees Method on Estimating the Car-Following Behavior
Sina Dabiri, Virginia Polytechnic Institute and State UniversityShow Abstract
Montasir Abbas, Virginia Polytechnic Institute and State University
Car-following models, as the essential part of traffic microscopic simulations, have been utilized to analyze and estimate the longitudinal drivers’ behavior since sixty years ago. The conventional car following models use mathematical formulas to replicate the human behavior in the car-following phenomenon. Incapability of these approaches to capturing the complex interactions between vehicles calls for deploying advanced learning frameworks to consider the more detailed behavior of drivers. In this study, we apply the Gradient Boosting of Regression Tree (GBRT) algorithm to the vehicle trajectory data sets, which have been collected through the Next Generation Simulation program, so as to develop a new car-following model. First, the regularization parameters of the proposed method are tuned using the cross-validation technique and the sensitivity analysis. Afterward, the prediction performance of the GBRT is compared to the world-famous GHR model, when both models have been trained on the same data sets. The estimation results of the models on the unseen records indicate the superiority of the GBRT algorithm in capturing the motion characteristics of two successive vehicles.
Assessment of Bridge Performance Through Machine Learning Algorithms: A Comparative Study
Fiorella Mete, Northwestern UniversityShow Abstract
Ying Chen, Northwestern University
Amanda Stathopoulos, Northwestern University
David Corr, Northwestern University
Modeling bridge performance is crucial to allocate resources in a cost-effective manner. Efficient statistical tools for bridge management rely on three key elements: (i) rich data sets; (ii) measurable metrics of bridge performance; (iii) efficient predictive models. In this paper, three machine learning algorithms are used to model the performance of a bridge subjected to a wide variety of truck loads: Multilinear Regression (MLR), Artificial Neural Network (ANN) and Regression Tree (RT). For this purpose, a modified measure of bridge deformation is adopted as a measure of bridge performance. This parameter, which combines strains recorded by a monitoring system and truck characteristics recorded by a weigh-in-motion system, is incorporated into each of the machine learning algorithms to quantify the effect of traffic conditions on the performance of the bridge. The relative benefits of each method are discussed and a strategy for model selection is provided along with a model performance comparison using cross-validation.
An Artificial Neural Network to Identify Pedestrians and Vehicles from Roadside 360-Degree Lidar Data
Junxuan Zhao, Texas Tech UniversityShow Abstract
Hao Xu, University of Nevada, Reno
Dayong Wu, Texas Tech University
Hongchao Liu, Texas Tech University
The high-resolution micro-level traffic data (HRMTD) of all road users is required for the full benefits of connected-vehicle technologies, especially in safety related applications. However, the data from traditional traffic sensors and limited connected vehicles cannot meet this data requirement. Cost-efficient roadside Light Detection and Ranging (LiDAR) sensors are identified as an effect solution to fill this data gap. One of the challenges in HRMTD extraction from roadside LiDAR data is to distinguish pedestrians and vehicles in all detection range. This paper proposed a new method to accurately distinguish pedestrians and vehicles based on artificial neural network (ANN). Performance of the trained ANN model was evaluated with field data, which confirms the effectiveness of the proposed method. The method introduced in this paper will be an important component of the roadside LiDAR processing system to provide accurate HRMTD to connected-vehicle applications. The HRMTD data will also greatly benefit the existing traffic safety engineering and traffic operation. The proposed method was developed for roadside LiDAR data processing. However, it can also be considered to process data from onboard sensing system of autonomous/semiautonomous vehicles.
Vehicle Reidentification in a Connected Vehicle Environment Using Machine Learning Algorithms
Zuoyu Miao, University of ArizonaShow Abstract
Larry Head, University of Arizona
Byungho Beak, University of Arizona
The deployment of Connected Vehicles (CV) is expected to become available for most American urban places in the next 10 to 20 years. The applications (e.g., mobility, safety and environmental) are constantly receiving equipped vehicles’ data. Even though the current ID protection mechanism assumes a vehicle’s ID changes every 5 minutes, the topic of re-matching vehicles is of great interests in privacy protection and performance measure research. This paper explores the possibility of re-matching connected vehicles’ ID using popular machine learning techniques, including Logistic Regression, Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Linear and Nonlinear Support Vector Machine (SVM) and Nearest Neighbor algorithms. An experiment is conducted using a microscopic traffic simulation model where data is collected through a Software-In-the-Loop technique. The best average mis-matching rate is as low as 14%. To assess potential factors’ effects on matching accuracy, a Poisson mixed regression model is analyzed under the Bayesian inference framework. The findings are: different matching algorithms vary in matching performance and the Linear SVM, the QDA and the LDA have the best accuracy results; traffic volume and market penetration rate have little impact on matching results; location and number of vehicles to be matched are considered significant. The results make the performance measure of future CV applications feasible and also suggest that data protection mechanisms should be investigated.
Integrated Cooperative Adaptive Cruise Control and Machine Learning Algorithms for Intelligent Vehicles Near an Off-Ramp
Changyin Dong, Southeast UniversityShow Abstract
Freeway diverge segment has significant impacts on the current traffic flow, and could affect the heterogeneous traffic flow consisting of manual and intelligent vehicles. The primary objective of this study is to evaluate how intelligent vehicles equipped with cooperative adaptive cruise control (CACC) improve freeway efficiency and safety at an off-ramp bottleneck. Based on NGSIM database, the lane-changing characteristics were learned from the ground-truth vehicle trajectory data utilizing randomized forest and back-propagation neural network algorithms. A microscopic simulation testbed was constructed, in which the realistic PATH CACC models and surrogate safety measures of the time exposed time-to-collision (TET) and time integrated time-to-collision (TIT) were used. The results showed that both CACC penetration and length of diverge influence areas exerted considerable influence on road capacity and traffic safety. Particularly, the capacity ascended to the peak after an initial decrease with the increase of CACC vehicles. The maximum capacity obtained in 100% CACC vehicle scenario was improved by over 53%, compared with 50% CACC penetration scenario. The proposed integration system with 100% CACC penetration can reduce the rear-end collision risks effectively, with the TIT and TET declined by 71.2%~97.8%. Moreover, the transport system with longer range of lane-changing area had better performance if all other parameters remain unchanged. Findings of this study can support freeway management and operations in the future.
Keeping Score: Incorporating Driver Behavior Scoring System with Connected Vehicles to Improve Traffic Service Quality
Ying Chen, Northwestern UniversityShow Abstract
Zihan Hong, Northwestern University
Yang Wu, Northwestern University
Hani Mahmassani, Northwestern University
In the era of intelligent transport, big data and connected systems, it is important to enable “smart” vehicles to identify and characterize individual drivers’ behavior rather than just collecting the mileage. This study aims to develop a driver scoring system to evaluate individual driving performance and improve the traffic condition and safety from the drivers’ perspective. The proposed scoring system adopts advanced data analytics techniques to extract, identify, characterize, and display driving habits and behavior patterns, including car-following and lane-changing behaviors, from vehicle trajectories. A safety score is developed by comparing a driver’s individual pattern to a standard “safe driver” pattern, defined by mining all drivers’ trajectories. The scores provide a basis for matching individual drivers in a connected environment, and suggesting to drivers an option for following another nearby “safe” driver. To evaluate the scoring system, a sample of trajectory data collected from anonymous drivers are used. In addition, the scoring system is integrated with a micro simulation tool with connected vehicle emulation capability. The results show that the car following recommendation system using the safety score improves overall performance of a connected traffic system beyond those attained through connectivity alone.
A Novel Approach to Missing Ramp Flow Imputation Using Machine Learning
Yuheng Kan, Zhejiang UniversityShow Abstract
It is a longstanding and tricky problem how to estimate/predict flows for any pair of on/off-ramps at which no measurement is available at all. In the context of machine learning, this problem is essentially a problem of transfer learning, i.e. no prior knowledge is available for the object of interest for estimation/prediction and hence based knowledge has to be drawn from objects of which some previous measurements are available. More specifically, this paper applies the random forest and gradient boosting methods to the problem and evaluates satisfactorily the methods based on field data collected from 9 pairs of on/off-ramps at the Shanghai city expressway.
Clustering Driver Behavior Using Dynamic Time Warping and Hidden Markov Model
Ying Yao, Beijing University of TechnologyShow Abstract
Xiaohua Zhao, Beijing University of Technology
Yiping Wu, Beijing University of Technology
Yunlong Zhang, Texas A&M University
Based on OBD and GPS installed in taxicabs, driver behavior data is collected. Left turn data on six similar curves is extracted, and speed, acceleration, yaw rate and sideslip angle of drivers in time series are selected as clustering indexes. Initial clustering is implemented by Dynamic Time Warping (DTW) and Hierarchical Clustering, and the clustering results are put into Hidden Markov Model (HMM) to iteratively optimize the results for achieving convergence. Driver behavior patterns over time while driving on the curves and the statistical characteristics of different groups are examined. All indexes including lateral vehicle control and longitudinal vehicle control have significant difference in different groups, indicates that the clustering method of DTW and HMM can effectively classify driver behavior. Finally, the driving behavior in different groups is further investigated and classified based on characteristics related to safe and ecological driving. This method can be applied by automobile insurance companies, and for the development of specific training courses for drivers to optimize their driving behavior.
A Novel Method of Mining Driving States via Latent Dirichlet Allocation Model
Zhijun Chen, Wuhan University of TechnologyShow Abstract
Hao Cai, Wuhan University of Technology
Yishi Zhang, Jinan University
Chaozhong Wu, Wuhan University of Technology
Bin Ran, University of Wisconsin, Madison
Automatic driving technology has attracted significant attention in artificial intelligence and data mining in the recent years. Mining driving states for accurately understanding driving behaviors is of significant importance for automatic vehicle. Although some methods for mining individualization driving have been developed, the application of these methods is still limited in practice because the latent driving states and structure driving behavior cannot be obtained. In this study, the objective is to mining the latent driving states and structure driving behavior for deeply understanding the individuation driving. The vehicle motion data are collected using on-vehicle detection sensors, and the driving behavior is successfully extracted by the proposed encode method. Latent Dirichlet Allocation (LDA) model is employed to discover driving states (topics) from individualization driving (documents) using driving behaviors (words). With the LDA model, the latent driving states and quantified structure of the driving behavior pattern can be discovered. In order to validate the performance and effectiveness of the proposed method, a typical unsupervised method k-means and the random method are selected as compared methods in our experiments. Experimental results show that the proposed method is effective can achieve better performance.
Performance Assessment of Urban Streets Adressing Improvement Issues for Automobile Mode of Transport
Suprava Jena, National Institute of Technology, RourkelaShow Abstract
Abhishek Chakraborty, IIT Kharagpur
Prasanta Bhuyan, National Institute of Technology, Rourkela
The present study focusses on modelling automobile drivers’ response pattern to assess the urban street service quality in developing countries. Several Quality of service attributes, affecting driver’s riding quality were investigated from 102 urban street segments under widely varying geometric and traffic conditions. From Pearson’s’ correlation analysis, total nine variables are found out to be significantly affecting drivers’ satisfaction level. Two novel Artificial Intelligence technique i.e. Artificial Neural Network (ANN) and Functionally Linked Artificial Neural Network (FLANN) are applied in this study to predict Automobile drivers’ level of satisfaction score ( ALOS_score). The prediction performance of developed models is assessed in terms various statistical parameters of a Modified Rank Index. Bayesian Regularization Neural Network (BRNN) algorithm has given the best fitted model in both training and testing data sets among the ANN models. However, application of FLANN model shows better prediction performance in the present context. Because, there exists no hidden layer and all the input layer neurons are directly linked with output layer neurons with lesser number of connections. Hence, it’s advantageous over ANN to reduce the accumulated error. The result shows that 73% of studied segments are offering service category ‘C’ or below. Sensitivity analyses have reported that Pavement condition is the most important variable with relative importance of 26.78% to influence the drivers’ riding quality. Similarly, other parameters were also ranked in decreasing order of their relative importance, which will help the highway authorities to prioritize budgets for future investments to improve provided service quality.
Gaussian Processes for Imputation of Missing Traffic Volume Data
Fabio Ramos, Federal University of Rio de JaneiroShow Abstract
Douglas Picciani, Federal University of Rio de Janeiro
Glaydston Ribeiro, Federal University of Rio de Janeiro
Heudson Mirandola, Federal University of Rio de Janeiro
Ivani Ivanova, Federal University of Rio de Janeiro
Saul Quadros, Federal University of Rio de Janeiro
Romulo Orrico Filho, Federal University of Rio de Janeiro
Leonardo Perim, DNIT-Brazil
Carlos Abramides, DNIT-Brazil
Despite the constant innovations of modern traffic counting devices, missing data is still a major concern for most of continuous traffic monitoring programs. In this work, it is proposed an algorithm for imputation of missing data obtained from continuous traffic counting devices using Gaussian Processes techniques for Machine Learning applied to time-series data. We present some applications of the method for the treatment of data from the Brazilian National Traffic Counting Plan (PNCT), under the responsibility of the Brazilian National Department of Transport Infrastructure (DNIT). The tests show that the PNCT data reconstructions obtained with the algorithm proposed in this work are robust and accurate. The method is flexible enough to consistently model widely varying traffic patterns, without loss of interpretability. Another important characteristic of the proposed method is the ability of generating confidence intervals for every imputed data. In the end of the work, we also discuss some possible future extensions for predicting traffic volumes, and to learn joint traffic patterns for different sensors.
Accelerating Stochastic Assessment of Postearthquake Transportation Network Connectivity via Machine Learning–Based Surrogates
Mohammad Amin Nabian, University of Illinois, Urbana ChampaignShow Abstract
Hadi Meidani, University of Illinois, Urbana Champaign
In earthquake prone areas, transportation networks are critical lifelines in providing access to the affected communities in the aftermath of an earthquake and support response and recovery efforts. Structural damages to transportation networks can disrupt such efforts and cause substantial socio-economic and physical losses. Therefore, evaluation of the transportation network reliability is essential for stakeholders and policy makers in order to facilitate optimal decision making for mitigation, preparedness, response, and recovery practices. Several research efforts have already addressed and quantified the impact of natural disasters on transportation networks, however, existing frameworks still suffer from high computational cost and thus are of limited applicability to large and complex networks. This paper presents a straightforward framework for accelerating simulation-based stochastic assessment of post-earthquake transportation network connectivity via machine-learning-based surrogates. The present framework enables fast risk assessment and real-time risk-informed decision making for large transportation networks. The proposed approach is applied to a simulation-based study of the two-terminal connectivity of a California transportation network subject to extreme probabilistic earthquake events. Different machine-learning-based surrogate models are considered in this study, based on support vector machine, logistic regression, k-nearest neighbors, and deep neural networks. A comprehensive comparison for the performance of these surrogate models is provided. Numerical results highlight the effectiveness of the use of deep neural networks in accelerating the transportation network two-terminal reliability analysis with surrogate accuracies of more than 99%.
A Novel Graph Partitioning Technique for High-Performance, Agent-Based Simulation of Fine-Resolution Travel Behavior
Husain Aziz, Oak Ridge National LaboratoryShow Abstract
Walking and bicyling are active transportation modes
that offer great potential to reduce carbon footprint from transportation sector
as well as pave the way to healthy living. Promoting active transportation
increases non-motorized vehicle miles traveled (VMT) which significantly
contributes to the reduction of traffic congestion and greenhouse gas emissions
from on-road vehicles. This research puts focus on the HPC implementation of the
agent-based model for New York City home-to-work commute trips and describes the
graph-based technique that has been applied for the execution and its
applicability in the context of travel behavior modeling. We built a HPC-based
ABM simulation framework to model travel related decisions at high spatial and
Processing Large-Scale Video Data to Support Transportation Safety, Planning, and Operations: A Flexible Approach to Data Storage and Integration
Venktesh Pandey, University of Texas, AustinShow Abstract
Weijia Xu, University of Texas, Austin
Lei Huang, University of Texas, Austin
Si Liu, University of Texas, Austin
Natalia Ruiz-Juri, University of Texas
Most traffic management centers in the United States use video feeds produced by monocular roadside cameras to monitor traffic conditions on a daily basis. This article presents a methodology to automatically process the data streams generated by such cameras in order to create data sets that can be queried and integrated with other data sources. Artificial intelligence libraries are used to recognize and detect moving objects from traffic videos, and information is stored into a structured queryable format that enables multiple analyses. We demonstrate the potential of the proposed framework by using automatically processed data in two distinct applications: traffic flow estimation and identification of pedestrian-vehicle conflicts. The vehicular counts produced by the analysis are over 90% accurate, and the technique is capable of identifying turning movements and close encounters between pedestrians and vehicles. Converting video data to structured data enables researchers to further explore applications of video feeds to support the evaluation and improvement of transportation systems. The data sets produced by this approach can be considered “raw” data, in the sense that they are not tied to any specific application. A variety of queries may be used to extract actionable information that complements the data available from other fixed and mobile sensors.
Selection of Highway Corridors Using Voronoi-Based Region Approximation and Artificial Intelligence Heuristics
Sushma Mb, Indian Institute of Technology, BombayShow Abstract
Sandeepan Roy, Indian Institute of Technology, Bombay
Avijit Maji, Indian Institute of Technology, Bombay
In planning process of highway development, it is important to explore the study area extensively and select the best alternative corridor. Selection of highway corridor represents a multicriteria decision process, which involves many factors (includes social, environmental, economic and engineering factors) and decision makers. These factors should be evaluated and compared for large number of alternative corridors. The model described in this paper provides an automated analytical approach to obtain a potential set of corridors. It is formulated based on cost function to seek an optimal set of corridors, and uses low-discrepancy based point sampling to extensively explore the search space. These sample points are converted into regions using Voronoi diagrams (VD). The VD uses information from boundary of the Delaunay triangulation (DT) generated around the sample points to estimate the cost of each region. Geographical information system (GIS) is used to extract the length and location dependent cost (environmental impact, right of way cost). The model uses Ant colony optimization (ACO) technique, one of the artificial intelligence-based optimization techniques, to come up with the optimum set of corridors. This study addresses the selection of appropriate set of corridors in a geographical region including rural or mixed rural-urban areas. The obtained corridors are compared for their quantitative and qualitative aspects like proximity to points of interest, environmental impact, geometrical aspects etc. A real world study area is used to illustrate the application of the model.
Keywords: Highway corridor, Voronoi, Low-Discrepancy, Ant colony, Artificial Intelligence
Predicting the Number of Uber Pickups by Deep Learning
Chao Wang, University of California, RiversideShow Abstract
Peng Hao, University of California, Riverside
Guoyuan Wu, University of California, Riverside
Xuewei Qi, University of California, Riverside
Matthew Barth, University of California
On-demand, app-based ride services like Uber and Lyft have become an important part of today's transportation system with its flexibility and quick responsiveness. Compared with traditional taxicabs, Uber-like taxis have loggers to monitor and record trip information such as pickup location and trip distance, which can be a valuable data source for knowledge discovering. Nowadays, a Real-time prediction for ride service demand (always reflected by the number of pickups) is increasingly crucial to improving the efficiency and sustainability of the urban transportation system. Newly aroused applications like ride sharing and autonomous mobility dispatching are based on solid demand predictions. In this paper, we propose a deep learning based approach to make dynamic predictions for Uber pickups using historical data. A Long Short-Term Memory (LSTM) Networks model is developed to learn the long-term Uber pickup patterns. With the experimental comparison of time-varying Poisson model and regression tree model, the results demonstrate the superior performance of our proposed deep learning model.
Scheduling for Timely Passenger Delivery in a Large-Scale Ridesharing System
Yang Zhang, Uber Technologies, Inc.Show Abstract
Husheng Li, University of Tennessee, Knoxville
Hairong Qi, University of Tennessee, Knoxville
Lee Han, University of Tennessee, Knoxville
Christopher Cherry, University of Tennessee, Knoxville
Taxi ride sharing is one of the most promising solutions to many urban transportation issues, such as traffic congestion, gas insufficiency, air pollution, limited parking space and unaffordable parking charge, taxi shortage in peak hours, etc. Despite the enormous demands of such service and its exciting social benefits, there is still a shortage of successful automated operations of ride sharing systems around the world. Two of the bottlenecks are: (1) on-time delivery is not guaranteed; (2) matching and scheduling drivers and passengers is a NP-hard problem, and optimization based models do not support real time scheduling on large scale systems.
This study tackles the challenge of timely delivery of passengers in a large scale ride sharing system, where there are hundreds and even thousands of passengers and drivers to be matched and scheduled. We first formulate it as a mixed linear integer programming problem, which obtains the theoretical optimum, but at an unacceptable runtime cost even for a small system. We then introduce our greedy agglomeration and Monte Carlo simulation based algorithm. The effectiveness and efficiency of the new algorithm are fully evaluated: (1) Comparison with solving optimization model is conducted on small ride sharing cases. The greedy agglomerative algorithm can always achieve the same optimal solutions that the optimization model offers, but is three orders of magnitude faster. (2) Case studies on large scale systems are also included to validate its performance. (3) The proposed greedy algorithm is straightforward for parallelization to utilize distributed computing resources. (4) Two important details are discussed: selection of the number of Monte Carlo simulations and proper calculation of delays in the greedy agglomeration step. We find out from experiments that the sufficient number of simulations to achieve a “sufficiently optimal solution” is linearly related to the product of the number of vehicles and the number of passengers. Experiments also show that enabling margins and counting early delivery as negative delay leads to more accurate solutions than counting delay only.