ABSTRACT. Spatiotemporal Traffic Data (STTD) measures the complex dynamical behaviors of the multiscale traffic system. Existing methods aim to reconstruct STTD using low-dimensional models. However, they are limited to data-specific dimensions or source-dependent patterns, restricting them from unifying representations. Here, we present a novel paradigm to address the STTD learning problem by parameterizing STTD as an implicit neural representation. To discern the underlying dynamics, we employ coordinate-based neural networks to directly map coordinates to traffic states. To unravel the entangled spatial-temporal interactions, we decompose the variability into separate processes. We further enable modeling in irregular spaces such as sensor graphs using spectral embedding. By integrating a meta-learning strategy, we decode instance-wise patterns to achieve efficient generalization across different instances. Through continuous representations, our approach enables the modeling of a variety of STTD with a unified input, thereby serving as a general learner of traffic dynamics. It also possesses properties such as implicit low-rankness, inherent smoothness, and reduced complexity, making it versatile for a wide range of practical applications. We validate its effectiveness through extensive experiments in real-world scenarios, showcasing applications from corridor to network scales. Empirical results not only indicate that our model has significant superiority over conventional low-rank models, but also highlight that the versatility of the methodology extends to different data domains, output resolutions, and network topologies. Comprehensive model analyzes provide further insight into the inductive bias of STTD. We anticipate that this pioneering modeling perspective could lay the foundation for universal representation of STTD in various real-world tasks.

ABSTRACT. Recent advancements in Connected Vehicle (CV) technology have prompted research on leveraging CV data for more effective traffic management. Despite the low penetration rate, such detailed CV data has demonstrated great potential in improving traffic signal performance. However, existing studies share a common shortcoming in that they all ignore traffic flow estimation errors in their modeling process, which is inevitable due to the sampling observation nature of CVs. This study proposes a CV data-driven robust optimization framework for traffic signal timing accounting for both traffic flow variability and estimation errors. First, we propose a novel data-driven uncertainty set of arrival rates based on the bounds information derived from CVs, which circumvents the error-prone arrival rate estimation process. Then, a generalized CV data-driven robust optimization model (CV-RO) that can be widely applied to various signalized intersection scenarios, including under-/over-saturated and fixed-/real-time, is formulated to explicitly handle arrival rate uncertainties. By means of the robust counterpart approach, this robust optimization problem can be equalized to a deterministic mixed-integer linear programming problem with an exact solution. The evaluation results highlight the superior performance of the CV-RO model compared to the deterministic model and traditional methods across various scenarios: different penetration rates, traffic demands, and control types. Notably, the CV-RO model demonstrates its excellence at lower CV penetration rates and in the presence of different traffic flow fluctuation levels, affirming its effectiveness and robustness.

ABSTRACT. This paper proposes a new spatial-temporal framework to predict the number of accidents in an urban area, by integrating multiple heterogeneous data sets such as road structure, intersection density, weather, historical collisions, traffic information and victim and party demographics data. Firstly we collect all data sets for the area of Los Angeles, which is our case study. Secondly, we propose a deep learning framework and efficient pipeline based on a three-dimensional convolutional autoencoder that determines the optimal size and sequential combinations of input for predicting the number of accidents to take place in a subarea at a given point in time. The framework is further improved based on victim and demographic party data, which indicates that the age of 23 is common among victims (mostly young men aged between 18-35 years old). Thirdly, we show that our proposed architecture outperforms other baseline models such as stacked denoising autoencoder, multilayer perceptron and convolutional long short-term memory network in terms of root mean square error and symmetrical mean absolute percentage error.

ABSTRACT. The generation of realistic samples of synthetic activity schedules is a critical component of both Activity-Based Models (ABMs) and simulation-based transport approaches such as MATSim. Conventional models decompose the scheduling process into series of discrete choices, applied sequentially. Realistic interactions between these choices are therefore limited. The combination of multiple interacting sub-models is also challenging to develop and use.

We demonstrate an alternative approach to activity sequence modeling using deep generative learning. Our model is able to rapidly synthesise activity sequences for new populations of socio-demographic attributes. This allows applications for (i) data anonymisation, (ii) diverse up-sampling, (iii) bias correction, and (iv) forecasting based on demographic shift. We comprehensively evaluate the quality of output activity schedules both individually and in aggregate.

We identify three primary benefits of our approach; (i) simplicity and speed, (ii) diversity in output sequences, and (iii) the potential for more realistic interaction of choice components. We have created the open-source Python software CAVEAT (https://github.com/fredshone/caveat) for the development and evaluation of generative activity models.

ABSTRACT. Technological advancements and interventions in the transportation sector play a crucial role in addressing climate change, given its major contribution to greenhouse gas emissions. The industry actively explores electrification, automation, and Intelligent Infrastructure to mitigate emissions. However, the successful design and implementation of these solutions require accurate and representative emission models.

The Motor Vehicle Emission Simulation (MOVES) serves as the gold standard emission software provided by the Environmental Protection Agency (EPA). Despite its prominence, using MOVES faces challenges, including a steep learning curve and technical complexities. This makes it cumbersome for macroscopic analysis and unsuitable for microscopic analyses like eco-driving, which demands emissions estimation for individual possible actions at each time step.

Efforts have attempted to address these needs by either pre-calculating emissions data or simplifying the MOVES framework. However, these MOVES variants either require extensive storage space and MOVES expertise, or sacrifice accuracy and comprehensiveness for accessibility and speed.

To address these issues, we present NeuralMOVES, a comprehensive family of high-performance and lightweight CO$_2$ emission models devised through reverse engineering MOVES and surrogate learning. Our models show a promising 6\% end-to-end error relative to MOVES, exhibit significant differences from alternative reduced-order models, and offer improved precision.

The implications of our work are twofold: our models simplify GHG emission evaluation in transportation-related analyses by providing a faster, programmatic alternative to MOVES and improve control-based approaches by offering microscopic and environment feature-rich models compared to alternative models.

ABSTRACT. In this paper, we extend the hybrid simulator from the individual to the household level by including a broader set of simulated demographic events affecting households and redefining a resampling procedure using the Gibbs Sampler. Usually, projection methods use historical demographic rates that may not account for sudden events like COVID-19, potentially hindering the accuracy of transportation models that rely on these projections. To test the resilience of projection methods to unforeseen events, we project synthetic samples from 2010 to 2021 using dynamic projection and a hybrid simulator. We test two scenarios based on pre-pandemic and post-pandemic demographic rates using Swiss Mobility and Transport Microcensus data. The results show that the hybrid simulator is more robust and less dependent on rates when it comes to unforeseen events than dynamic projection as it includes an intermediate resampling update that helps reduce the errors of dynamic projection.

ABSTRACT. Modeling pedestrian entrance scenarios is a central focus in the field of pedestrian dynamics, yet existing models, rooted in physics, have limitations when it comes to incorporating psychological aspects of individual behavior. Despite prior efforts to integrate certain psychological concepts, this interdisciplinary perspective is relatively new, and there is room for further exploration. This study aims to initiate a discourse on the integration of motivation into pedestrian models. Motivation is believed to be one of the most apparent psychological drivers of movement behavior in pedestrian environments, capable of significantly influencing crowd dynamics. While previous approaches have often employed a simplified binary categorization of motivation, classifying agents as either highly motivated or lowly motivated, this simplification, while useful in many contexts, fails to capture the complexity of motivation, which is influenced by a multitude of intrinsic and environmental factors. We introduce two critical dimensions of motivation: heterogeneity (variations in individual motivation levels within the crowd) and dynamism (fluctuations in motivation levels during goal pursuit) to establish a foundation for modeling motivation in entrance scenarios. Incorporating these dimensions alongside related psychological motivation theories, we present a simple velocity-based model.

ABSTRACT. With increasing deployment of automated vehicles (AVs), it is crucial to understand how they will interact with human-driven vehicles (HDVs). We investigated the gap acceptance behavior of HDVs in mixed traffic at a priority T-intersection. We adopted a driving simulator, in which drivers drove four scenarios, that varied on the driving style (less defensive, more defensive) of AVs and on their recognizability (distinguishable or not from HDVs). Then, using the collected data, we estimated gap acceptance models, which we then implemented in the SUMO microscopic traffic simulation platform, in which we built a T-intersection network. We ran simulation runs differing in AV driving style, recognizability, penetration rate (0-75% in 25% increments), and whether HDV behavioral adaptation was considered. HDVs were present on the minor and major road, while AVs were only present on the major road. Vehicle delays and queue length were measured as performance indicators. We found that the delay for minor road vehicles increased with higher penetrate rate of AVs. Moreover, the minor road delays were larger when AVs were recognizable and less defensive. We also found that if behavioral adaptation was ignored, then the delay for minor road vehicles would be underestimated by up to 75%. In conclusion, we observed HDVs’ behavioral adaptation in gap acceptance behavior in mixed traffic. The recognizability of AVs, their driving style, and their penetration rate affected HDV behavior. If we want to predict and assess efficiency of mixed traffic, then considering behavioral adaptation is important.

ABSTRACT. The development of autonomous driving technology offers new opportunities for traffic management, with connected and automated vehicles (CAVs) expected to liberate drivers for autonomous cruising. However, the continued need for short-distance travel by foot or bicycle will result in traffic flows comprising both vulnerable road users (VRUs) and CAVs. Managing their interactions, particularly at accident-prone automated unsignalized intersections, poses a significant challenge: developing a conflict-free and efficient operational strategy in such multi-player, dynamic, and continuous interactions. In this work, responding to these practical needs and research challenges, we propose a dynamic interaction-based cooperative decision-making approach for heterogeneous traffic participants (DI-CDM). We introduce a differential game framework aimed at analyzing interactions between VRUs and CAVs, to ensure the safe and efficient operation of unsignalized intersections. The results demonstrate that the proposed DI-DCDM exhibits promising performance in managing interactions among heterogeneous traffic participants and generating conflict-free trajectories.

ABSTRACT. While studying the interactions between pedestrians and urban traffic remains important, there's a growing urgency to investigate pedestrian behaviour within the framework of automated vehicles (AVs). This study uses virtual reality (VR) to simulate two urban mid-block environments, one in downtown Toronto, Canada and the other in central Newcastle, UK, to investigate the crossing behaviour of 428 participants (9,092 observations). To the best of our knowledge, this is the largest study of such kind conducted on two different continents with different traffic rules and walking/driving norms. The research questions addressed in this paper, in the context of unmarked mid-block crossing, are: (a) Do various vehicle types, i.e., normal vehicles and AVs, impact pedestrian behaviour? (b) How do other pedestrians influence one's crossing behaviour? (c) How do traffic characteristics, road type, and environmental characteristics impact pedestrian behaviour? (d) What is the influence of demographics on pedestrian behaviour? and, Are there differences in pedestrian behaviour in different countries? We use advanced econometrics and data driven machine learning to answer these questions.

ABSTRACT. In the field of autonomous driving, precise evaluation of drivers' cognitive states through physiological indicators is essential for advancing safety and vehicle interaction. This investigation assesses the association between eye movement patterns and cognitive states across varied driving scenarios, including Non-Driving Related Tasks (NDRT), light intensity variations, and different driving modes. Eye movement data were collected from experienced drivers during simulations, utilizing Tobii Pro Glasses 3. The data analysis was conducted using SHapley Additive exPlanations (SHAP) to interpret the predictive power of eye movement features within a Light Gradient Boosting Machine (LightGBM) model comprising nine distinct features. The findings highlight that fixation-related features significantly influence cognitive load states. Further analysis involved clustering fixation data into seven areas, which served as inputs for a Markov chain model to simulate transition patterns among cognitive states. The study revealed that increased cognitive load correlates with more frequent checks of the dashboard, whereas lower cognitive load corresponds with increased focus on the road ahead. Additionally, the effects of NDRT and light intensity on eye movement patterns were more pronounced than those of the driving mode. These insights underline the potential of utilizing eye movement metrics to enhance safety features in autonomous vehicles, reflecting the unconscious behaviors of drivers under different cognitive loads.

ABSTRACT. Heterogeneity has been a longstanding characteristic of human-driven vehicle traffic. It is set to be compounded by the growing adoption of automated vehicles (AVs) and the emergence of mixed traffic. The behaviors exhibited by AVs differ among themselves and notably with human driven vehicles (HDVs), highlighting the complex and heterogeneous nature of mixed traffic.

Central to our work here is the challenge of data-driven learning under such heterogeneity. Specifically we recognize that some methods of aggregating data across diverse traffic agents (e.g., different human drivers) to develop predictive models (e.g., driving models) encounter significant limitations. Such methods often fail to accommodate the distinct behaviors and preferences of individual agents. Misalignment between how traffic agents desire to behave versus how we model them to can mislead both data-driven predictions and traffic-scientists alike.

We propose a framework that employs a personalized modeling\footnote{personalization refers to models that adapt their learning to agents' behavior and preferences} approach to learn models for traffic agents\footnote{by traffic agents we mainly refer to HDVs and AVs, yet we leave the notation general as the modeling framework is not restricted to particular agents. For instance, bicyclist can be seen as traffic agents.}. This approach is motivated by (i) the heterogeneous nature of driving data and challenges it poses on data-driven learning, and (ii) the distinctive requirements of mixed-traffic environments, which demand effective human-machine coexistence and behavioral alignment.

Personalization offers a valuable opportunity to identify and utilize behavioral differences among traffic agents, and leverage them for improved downstream analytics (i.e., operation management).

ABSTRACT. This study aims to illustrate the critical role of user psychological factors in travel choice and travel behaviour via social media data and natural language progressing methodologies. By integrating social media with survey data to obtain more comprehensive psychological features including dynamic user attitude and user concerns through sentiment analysis and dynamic topic modelling, we are able to demonstrate how these features help in identifying underlying causes and key determinants of both consistency and inconsistency travel behaviour within neighborhood level and different periods. We implement a case study in New York City via Twitter data to capture transit-related psychological factors on five travel modes (cycling, driving, subway, taxi and ride-hailing, walking) from 2019 to 2022. Results show the advanced ability of social media data in capturing dynamic travel-related user attitudes and concerns and further support travel choice dissonance identification and dynamic travel behaviour explanation among five travel modes. Consequently, the outcome efficiently reflects various social needs, identifies priority areas and provides valuable suggestions for policy-makers and planners to develop more targeted improvement strategies.

ABSTRACT. The causal graph is an indispensable part of causal analysis, which plays a considerable role in congestion and accident investigation, as it provides an intuitive form of causal correlations in traffic networks and the propagation process of causal impacts. Previous studies have focused on generating the causal graph using information theory and model-based approaches, which ignore region-wise mutual information and network topology, hindering the comprehensive understanding of the causal relations. To address such limitations, this work provides a new perspective to achieve region-wise causal inference of complex traffic networks, enhancing verifiability and interpretability. Specifically, the trainable causal entropy algorithm is proposed to obtain the initial causal matrix, considering the topology structure of the network. A causal-integrated Graph Convolutional Network (GCN) with an attention mechanism is then designed to capture region-wise causal effects by information aggregation. The initial causal graph is iteratively refined within an end-to-end framework. The effectiveness of the proposed method in capturing causal relations is evaluated on a real-world dataset through the improvement observed in traffic forecasting accuracy. Additionally, the distinct visualizations of the symmetric topology graph and the asymmetric causal graph reveal the complex interplay of traffic dynamics, highlighting the region-wise causal impacts that shape the underlying relations within the traffic network.

ABSTRACT. This paper aims to understand how physics-based and AI models interact and supplement each other in modeling car-following (CF) behavior. Physics-based CF models are grounded in robust theoretical frameworks, offering a level of interpretability and trust in their output. Nonetheless, achieving a simplified theoretical exploration comes at the expense of modeling complex dynamics and interactions in generalized traffic environments. This makes adopting traditional CF models to automated vehicles (AVs) control more challenging, given the crucial need to realize generalizability. Conversely, the emerging AI models (i.e., deep learning (DL)) offer unique advantages in learning and adapting from data, particularly capturing non-linear complex relationships. Yet, this requires a vast amount of high-quality training data that is often limited or unavailable. More importantly, the AI methods lead to models with few to no interpretable insights, impairing traffic-level understanding. Consequently, we see a shift in adopting AI-based models to supplement traditional CF models – where they might fail – giving rise to physics-informed artificial intelligence (AI) models, most famously physics-informed neural networks (PINN) (Mo et al., 2021, Cuomo et al., 2022). The rigorous analysis of their approximation efficiency remains unexplored. In this work, we want to understand the behavior of these physics-informed AI models. This entails analyzing how AI models interact with traditional CF models and what expectations we have on the synergy between the two. Incorporating physics-informed AI into CF models brings multiple dimensions of analysis: predictive accuracy, data needs, convergence, interpretability, etc. All these play a role in physics-AI synergy. However, in this abstract we focus on the accuracy metric. Our interest here is to investigate how the underlying physical features in different CF models across different families synergies with AI models. We do so through theoretical error bound modeling and sensitivity-based simulations.

ABSTRACT. Recent studies have significantly advanced our understanding of the fractal nature of traffic flow, suggesting that it can be characterized within the Kardar-Parisi-Zhang (KPZ) universality class. This framework proposes that traffic dynamic is governed by KPZ exponents: α = 1/2 for traffic jam size distribution, reflecting spatial roughness scaling; β = 1/3 for the scaling of traffic fluctuations over time, indicating temporal growth; and z = 3/2 for the relationship between total delay and lane-mile length, indicating space-to-time scaling. This study builds upon these theoretical insights by empirically validating these critical scaling laws using real-world traffic data from the I-24 MOTION dataset. By aligning empirical observations with theoretical predictions, this research aims to substantiate the applicability of the KPZ class to traffic flow. This study opens the door to potential congestion mitigation strategies that could control KPZ scaling exponents to maintain the traffic system in a sub-critical state.

ABSTRACT. In the extended abstract we present a partial differential equation (PDE) model consisting of three parts, a crowd flow dynamics component, an epidemics spreading model, and an equation that provides the velocity field due to ventilation. Using a specific numerical implementation, we present preliminary tests, in which the crowd is moving towards an exit that is assumed to be the whole right boundary of the computational domain. The numerical results illustrate the behavior of epidemics transport under movements of pedestrians in a closed area, in the presence of ventilation.

ABSTRACT. In this paper, we question the assumption of “small” disturbances in the study of string stability of car-following models and ACC systems, which is at the basis of analytical approaches to linear analysis. This assumption is in fact unrealistic in real traffic operations, where controllers are typically nonlinear and disturbances are “large”; see sharp decelerations. Therefore, evidence of the inaccuracy of existing analytical methods for nonlinear analysis and their inapplicability to a large class of car-following models is first presented. A simulation framework is then proposed and applied to identify the limits of linear analysis and to disentangle the occurrence of rear-end collisions from that of string instability. Metastability thresholds, which define how large a perturbation must be for a linearly stable model to become unstable, are calculated at different equilibrium states for the IDM and Gipps’ model. Analytical conditions for metastability as a function of model parameters are also derived. The analysis showed that no collisions due to the amplification of a disturbance occurred for both metastable and linearly unstable parameters, even when the deceleration output of the two models was constrained not to exceed normally achievable values. However, the introduction of a perception delay made it possible to obtain vehicle collisions already in the metastable region. Finally, the study addresses a major criticism of Gipps’ model. When string stability is measured by the L_∞ norm, Gipps’ model becomes unstable for decreasing values of the equilibrium speed – in agreement with real traffic observations – both in the metastable region and in the linear instability region.

ABSTRACT. Network signal control is an effective way to deal with nonlinear congestion dynamics in urban networks. However, most network signal control methods ignore queue spillback. Few studies that considered queue dynamics assume homogeneous traffic states across different movement directions in one road segment, resulting in the underrepresentation of spillback. This study proposes a novel hierarchical model predictive control (MPC) approach that utilizes the link transmission model (LTM) with queue transmission at both road segment and turn levels to optimize signal timing for urban networks. At the network level, the controller employs a state-of-the-art LTM that can describe segment-level flow dynamics and uses quadratic programming to determine the reference effective fractions of green time for network throughput maximization. The local layer builds on a newly developed LTM with turn-level queue transmission and uses nonlinear programming to track the reference of green time to ensure that the optimization by the network layer is realized at the individual intersections while eliminating potential spillbacks. Simulation experiments on an arterial are conducted to verify the performance of the proposed approach. The results reveal that the proposed MPC leads to significant improvements in traffic performance in terms of throughput and queue length, especially in oversaturated conditions.

ABSTRACT. In this work, we aim to evaluate three representative ConOps in drone delivery services. Concept A utilizes pre-planned routes with no intersections, where trajectories are confined within these routes. Concept B also employs pre-planned routes but allows intersections, with trajectories still restricted to these routes. Concept C does not pre-plan routes but generates 4D trajectories that could be spatially or temporally separated. The three concepts vary in terms of flexibility, infrastructure and control requirements, and monitoring challenges. In this study, we aim to quantify these differences by considering specific built environments and order characteristics.

ABSTRACT. While Urban Air Mobility (UAM) holds promise for improving urban travel, its integration with existing Air Traffic Management (ATM) systems poses significant challenges. As UAM technology evolves and more VTOLs enter shared airspace, ensuring interoperability with ATM becomes increasingly complex (Levitt et al., 2023). One cautious approach is to design exclusion-area airspace for UAM, simplifying management but potentially limiting operational capabilities. Vascik et al. (2020) emphasized the need for balance, where excluding UAM operations from airspace used by ATC limits available airspace, especially at higher altitudes. Similarly, simulations in the Dallas-Fort Worth area reveal challenges in achieving separation between UAM and conventional aircraft, even with well-designed corridors (Lee et al., 2022). As operational tempo increases and UAM technology progresses, there will be an increasing requirement for coordination between UAM and ATM operations. This coordination becomes crucial as automation capabilities are developed and sufficiently validated beyond UML-4 (Levitt et al., 2023). Despite its promise, relatively few studies have examined the feasibility of integrating UAM operations within airspace, especially in regions with complex flight patterns. One key aspect of safely integrating UAM with ATM is an accurate prediction of non-UAM conventional aircraft trajectories to mitigate collision risks and to ensure the necessary separation between UAM and conventional aircraft. While conventional aircraft typically follow predefined routes, uncertainties can arise especially in terminal airspace due to heavy traffic, aircraft dynamics, environmental conditions, or human factors. To this end, we propose a framework for assessing the feasibility of UAM routes near terminal airspace by leveraging probabilistic aircraft trajectory prediction. Our framework utilizes short-term predicted trajectory distribution information generated by a flow-based deep-generative model, called Normalizing Flows. This information is used to evaluate UAM routes intersecting ATM airspace at various altitudes, with UAM aircraft to decelerate by anticipating encounters with conventional aircraft in the short term.

ABSTRACT. Unmanned Aerial Vehicles (UAVs), referred to as drones, can organize themselves into swarms, fostering collaboration and efficiency in sensor data collection within Smart Cities. With their mobility, autonomy, and diverse sensors, drones have been widely used in the transportation systems. For instance, drones can be used for an accurate monitoring of traffic to detect traffic congestion at early stage. This allows traffic operators to apply mitigation actions that decrease the carbon footprint of a sector with one of the highest carbon emissions worldwide.

Utilizing multiple low-cost drones over a wide sensing area offers a flexible alternative to single high-profile drones. They complete sensing missions in parallel, benefiting from shorter recharging times. This requires coordinated actions with autonomy and computational intelligence. Recent advancements in decentralized optimization and multi-agent learning algorithms offer scalability and efficiency while maintaining privacy and autonomy. However, developing, testing, and evaluating such solutions is complex. Simulation environments simplify studying swarm intelligence and collision avoidance algorithms by reducing complexity and environmental variables. In contrast, real-world drone experiments indoors and even outdoors enhance realism and external validity.

To bridge this gap, this paper introduces a testbed to study distributed sensing problems of drones, such as such as energy consumption, charging control, navigation and collision avoidance. This testbed sets a stepping stone to emulate, within small laboratory spaces, large sensing areas of interest originated from empirical data and simulation models. As a proof-of-concept, a multi-agent collective learning approach is applied to this testbed to coordinate and optimize in a fully decentralized way the navigation and sensing of drones. Furthermore, a potential field collision avoidance method is applied to predict the fields of collisions and finds the optimal flying trajectories of drones to mitigate the risk of collisions and sensing inefficiency. Extensive experimentation using real-world data in traffic monitoring in Athens city validates the efficiency in traffic vehicle observation, demonstrating the capacity of the testbed to move complex swarm intelligence and collision avoidance algorithms for drones to real-world.

ABSTRACT. Gaussian process model UAV trajectory planning traffic state estimation uncertainty quantification uncertainty propagation

ABSTRACT. This paper addresses the challenge of estimating dynamic OD matrices efficiently by drawing inspiration from video summarization techniques. Like summarizing video sequences into key shots, our proposed framework focuses on key intervals, reducing computational requirements while maintaining accuracy. The proposed framework involves Shot Boundary Detection (SBD) and Interpolation components. SBD identifies key intervals, crucial for accurate estimation. We propose a new SBD algorithm based on individual demand and interpolation feedback. Interpolation fills gaps between key intervals using techniques like linear and spline interpolation. Moreover, Split Simulations is also introduced to reduce lower-level bi-level adjustment complexity by capturing assignment matrices only for key intervals With the case study on a real network from Logan, Queensland, Australia, the paper provides justification of gradual travel demand evolution, requiring estimation for a subset of time intervals (key intervals) for accurate OD estimation. It is demonstrated the reduced runtime enables more iterations in bi-level OD adjustment, enhancing convergence. The performance of the proposed KFI method is compared against a state-of-the-art bi-level adjustment framework using a generalized least squares formulation. Results indicate a 4.7% reduction in runtime with our framework and an 80.15% reduction with split simulations compared to traditional methods. Additionally, performance evaluation based on the GEH statistic validates our approach's accuracy, surpassing industry standards.

ABSTRACT. In response to the rapidly evolving urban landscape, there is a growing demand to enhance traditional Origin-Destination Matrices Estimation (ODME) models with new data sources that offer broader perspectives. Emerging technologies and crowd-sourced data in particular, can provide promising avenues for collecting high-resolution data on destination activities, reflecting real mobility patterns. Previous research by Castiglione et al. (2024) investigates trip motivations and the varying travel flexibility associated with different activities, leveraging real-world crowd-sourced data like Floating Car Data (FCD) and Google Popular Times (GPT).

To address the need to integrate these insights into ODME models, this paper introduces the Flex-GLS approach, an extension and generalization of the GLS model. Flex-GLS accounts for multiple demand components characterized by spatio-temporal flexibility metrics derived from crowd-sourced data. It aims to offer a more precise representation of travel demand by integrating both temporal and spatial flexibility dimensions. Benchmarking against the traditional GLS model highlights the potential of Flex-GLS in enhancing ODME accuracy, providing valuable insights into urban travel dynamics.

ABSTRACT. Origin-destination (OD) demand matrices are crucial for transit agencies to design and operate transit systems. This paper presents a novel temporal Bayesian model designed to estimate transit OD matrices at the individual bus-journey level from boarding/alighting counts at bus stops. Our approach begins by modeling the number of alighting passengers at subsequent bus stops, given a boarding stop, through a multinomial distribution parameterized by alighting probabilities. Given the large scale of the problem, we generate alighting probabilities with a latent variable matrix and factorize it into a mapping matrix and a temporal matrix, thereby substantially reducing the number of parameters. To further encode a temporally-smooth structure in the parameters, we impose a Gaussian process prior on the columns of the temporal factor matrix. For model inference, we develop a two-stage algorithm with the Markov chain Monte Carlo (MCMC) method. In the first stage, latent OD matrices are sampled conditionally on model parameters using a Metropolis-Hastings sampling algorithm with a Markov model-based proposal distribution. In the second stage, we sample model parameters conditional on latent OD matrices using slice and elliptical slice sampling algorithms. We evaluate the proposed model using real-world data collected from three bus routes with varying numbers of stops, and the results demonstrate that our model achieves accurate posterior mean estimation and outperforms the widely used iterative proportional fitting (IPF) method. Additionally, our model can provide uncertainty quantification for OD demand matrices, thus benefiting many downstream planning/operational tasks that require robust decisions.

ABSTRACT. Demand-responsive multimodal transit offers opportunities to complement existing public transport systems and provide an overall better service level to passengers while at the same time making better use of the resources. This study optimizes the capacity of such system by strategically sizing the required fleet and allocating it to the operating services. We formulate a two-stage stochastic optimization model that plans the transit system and the required fleet in the first stage, and optimizes the demand-responsive operations in the second stage. We develop a decomposition-based method that exploits the network-based formulation of the second stage, allowing us to solve practical instances. Preliminary results from a case study in the city of Zurich show that designing a public transport system together with demand-responsive mobility systems can benefit both transport operators and passengers. By allocating the system capacity more efficiently, operators reduce operational costs while maintaining or improving the travel experience for passengers.

ABSTRACT. Empirical evidence suggests that couriers often deviate from pre-planned delivery routes, responding to practical realities that routing algorithms may overlook, such as dynamic traffic conditions or recipient availability. This study explores the reasons behind delivery drivers’ deviations from assigned routes, particularly how their accumulated experiences inform their navigation choices. Using data from the 2021 Amazon Last-mile Research Challenge, which labels courier-performed routes based on practical considerations like productivity and customer satisfaction, we demonstrate that the quality of a delivery route is influenced by more than just travel time. Factors like turn sharpness, backtracking distance, and neighbourhood visit timing also have a statistically significant impact on route quality. This study trains an energy-based model to predict the likelihood of a route being of high quality when executed in the field. We integrate the model’s insights into an insertion heuristic to generate high-quality last-mile delivery routes.

ABSTRACT. Wider adoption of Electric Vehicles (EV) promises major benefits in terms of reductions in CO2 and air pollution. However, access to charging infrastructure presents a major barrier to mass adoption and continued use of EVs. Charging scheduling has attracted much attention in the literature as it helps to effectively manage limited charging infrastructure and power grid capacity, thereby relieving EV users' charging anxiety. However, these studies treat the demand for vehicle charging as direct, therefore omitting the inherent link with daily scheduling behavior. This study aims to fill the gaps by incorporating EV charging into activity scheduling to capture the inherent links and tradeoffs between charging and activity scheduling. This study introduces a new activity-based framework to estimate charging behavior. Having such a model allows us to realistically understand charging and travel behaviors with a high level of temporal and spatial resolution. It also enables the evaluation of the effects of alternative management strategies (e.g., Time-of-Use tariff, infrastructure build-out) on individual charging and travel behavior.

ABSTRACT. This extended abstract presents a real-time shared ride multi-depot DAR service where vehicle speeds are variable with time. Accumulation based NMFDs in statically partitioned regions are used to estimate vehicle speeds at each time step. The DARP is formulated as a multi-objective MINLP that considers the conflicting priorities of maximising the operator's total profit and minimising the users' total delay and number of rejections. Once a request has been accepted it may not later be rejected. Requests can be reassigned between vehicles until a vehicle is en-route to their pick-up or drop-off node. Initial computational experiments were conducted using Gurobi and found that the model can provide real-time solutions while considering dynamic traffic states and pre-existing requests. This extended abstract only considers vehicles with internal combustion engines, however the full paper will consider a service fleet of BEVs. The DARP will be edited to include charging facility selection and scheduling. The charging facilities will be capacitated and a queuing model will be used to estimate waiting times at each facility considering stochastic charging demand from private BEVs. The key contribution of the final paper is the joint consideration of the regional dynamic traffic model and stochastic charge queuing model during optimisation.

ABSTRACT. In this study, we propose a novel continuum approximation approach to optimally plan urban public charging infrastructure for electric vehicles, considering spatial heterogeneity, serviceability, and region-wide equity. We analytically estimate waiting and travel times to charge with queueing theory and evaluate serviceability over a planning region with spatial heterogeneity. The analytical model explicitly identifies three groups of factors to determine serviceability: (i) planning factors, including the spatial heterogeneity of station density and number of chargers per station; (ii) operational factors, such as station assignment rules; and (iii) exogenous factors, including charging demand, roadway network, and traffic conditions. We validate our planning framework via a numerical example within the urban network and a case study focusing on New York City. We find that solution varies with spatial heterogeneity, high-demand areas require more stations and chargers, while low-demand areas need some for equity concerns.

ABSTRACT. Electric vehicles (EVs) and the infrastructure of electric vehicle charging stations (EVCS) are emerging as essential components of sustainable energy systems. In this context, we introduce an innovative approach that utilizes aggregated EVCS to participate in the auxiliary market, thereby providing grid-balancing services. Our model continuously monitors changes in EV state-of-charge (SoC) across both time and space, taking into account various factors, including driver behavior, current SoC levels, and the associated charging/discharging costs and benefits. This approach will enable charging station operators (CSO), in collaboration with aggregators, actively engage in the frequency containment reserves (FCR) market. We introduce an optimization framework in conjunction with this EV model. For establishing pricing policies with the twin aims of maximizing profits for aggregators and charging station operators (CSOs), while also minimizing energy charging expenses for EV users. Our findings underscore the effectiveness of this pricing strategy in achieving these dual objectives, as demonstrated through realistic simulations integrating the EV mobility and the Electricity FCR market.

ABSTRACT. This research investigates the use of large language models and machine learning techniques to identify the primary triggers for air traffic flow management regulations. The study focuses on textual remarks made by flow managers who implemented these regulations. The investigation takes a concrete form by using weather-related regulations with the referenced location being an aerodrome. Specifically, a large language model is asked to assign each of these regulations to a specific group, or cluster, based on the remark made by the flow manager, where each cluster represents a particular kind of weather disruption. These clusters then act as labels for the dataset, and each regulation is combined with the weather conditions observed during its implementation. This labelled dataset is then used to train a tree-based classifier using supervised learning. This two-step methodology enables the identification of the most likely precise trigger for each regulation, such as low visibility, snow, strong winds, etc. based solely on observed weather conditions. The clusters identified by the large language model are also compared with those discovered in previous research using self-learning and supervised clustering. Nevertheless, the practical applications of this method go far beyond the classification of weather-related regulations. This approach could be used in post-operational analysis to identify the primary triggers of any type of regulation - not just weather-related. Furthermore, it enables the analysis and classification of other types of text, such as notices to airmen, further broadening its potential use cases. This paper showcases the versatility and broad application of large language models in the field of air transportation.

ABSTRACT. Accurate detecting of road markings is indispensable for autonomous vehicles (AVs) to navigate safely and efficiently. However, current autonomous driving systems still face challenges in perceiving static and dynamic road information leading to inaccurate trajectory and path planning processes, especially in complex driving scenarios. To address these challenges, recent years have seen a predominant reliance on deep learning and attention mechanisms for road marking detection (Jayasinghe et al., 2022). However, persistent challenges such as high computational demands, extensive dataset requirements for training, and manual data annotation issues necessitate exploring alternative approaches. Recent advancements in Large Language Models (LLMs) and Vision Language Models (VLMs) present the possibility of addressing these challenges (Sha et al., 2023). While LLMs have demonstrated remarkable capabilities in reasoning tasks, VLMs offer a distinct advantage by integrating visual and textual information, thereby providing a more comprehensive understanding of complex scenarios (Zhang et al., 2024). Consequently, VLMs are instrumental in interpreting complex driving scenarios and enhance autonomous vehicles’ decision-making processes in real-time (Sha et al., 2023). As a result, this paper proposes a new road marking detection model based on VLM to address the aforementioned challenges. This model leverages both image and label information of road markings, achieving optimal detection results even with relatively small datasets. Furthermore, in conventional autonomous driving, perception and decision-making are two relatively independent processes, and to fully utilise the results of perception to assist the decision-making process, a comprehensive end-to-end pipeline is established that uses GPT-4 (Generative Pre-trained Transformer), an LLM-based language model, for question answering (QA) of road scenarios based on model’s output. This pipeline enhances the autonomous driving system's ability to perceive road surfaces and aids in the decision-making process while driving. The key contributions of this paper are outlined as follows: (1) Development of a new road marking detection model based on VLM, offering reduced training costs compared to traditional methods, (2) Construction of an end-to-end pipeline to enhance the autonomous driving system’s comprehension of road information and generation of driving prompts to support decision-making processes.

ABSTRACT. This study aimed to evaluate the performance of VideoChatGPT, a large video language model, for automatic crash detection and narration from traffic camera videos. For this purpose, a dataset of 500 crash videos was processed, ranging from 2 to 40 minutes. The results showed that VideoChatGPT accurately detected crash times in 2-minute videos, with 72\% accuracy within 4 seconds. However, it struggled with longer videos until a sub-sampling strategy was applied. Additionally, the analysis of 2-minute crash narratives revealed that the model had high accuracy (73.5\%) in detecting vehicle dynamics but very low accuracy (<5\%) in identifying specific collision types such as head-on crashes and rollovers. Environmental details were described accurately 92\% of the time, while contextual details such as road type had 67\% accuracy. However, the model lacked accuracy in causal reasoning and overall description precision with only 54\% accuracy. These findings highlight the current capabilities and limitations of VideoChatGPT, which requires further training on diverse traffic data to improve its automatic video understanding for traffic incident management.

ABSTRACT. Spatiotemporal traffic data can be modeled as a multivariate time series. A crucial task in spatiotemporal analysis is to identify and detect anomalous observations and events from the signals with complex spatial and temporal dependencies. Robust Principal Component Analysis (RPCA) is widely used tool for anomaly detection. However, the default RPCA purely relies on the global low-rank assumption while ignoring the local temporal correlations. In light of this, this study proposes a Hankel tensor version of RPCA for anomaly detection in traffic data. We treat the spatiotemporal traffic data as a matrix (location by time) and decompose the corrupted raw matrix into a low-rank Hankel tensor and a sparse matrix. By leveraging the Hankel constraint on the temporal dimension, the model can simultaneously capture the global and local spatiotemporal correlations and exhibit more robust performance. We formulate the problem as an optimization problem and develop an efficient solution algorithm based on the Alternating Direction Method of Multipliers (ADMM). Despite there are three hyper-parameters in the model, they are easy to set in practice. We evaluate the proposed method by metro passenger flow collected from Guangzhou, China, and the results demonstrate the efficiency of anomaly detection.

ABSTRACT. This article introduces a method to automatically detect, track and stitch the stop-and-go wave fronts and tails in space and time for large-scale vehicle trajectory, enabling the analysis of stop-and-go wave characteristics at scale. The contributions of this paper are outlined as follows: (1) we proposed a method capable of detecting, tracking and stitching both the wave fronts and tails in stop-and-go traffic. (2) we implemented the proposed methods in Python to support reproducible research, data and tools will be provided at i24motion.org. (3) we applied our method to a large-scale dataset, demonstrating the complex phenomena of wave generation, propagation, merge, bifurcation, and dissipation.

ABSTRACT. While data is ubiquitous and transforming almost every aspect of transportation, the wide application of data-driven methods also reveal vulnerability of such methods when facing data poisoning attacks. This paper proposes a defense framework against data poisoning attacks using secure data from the infrastructure. We show the three key steps of this infrastructure-enabled defense (IED) framework, including the selection of the secure data, attack detection, and attack mitigation. The IED method is tested using data poisoning attacks on queue length estimation using mobile sensing data. Performance of IED is reported with comparisons with a few benchmark defense methods.

ABSTRACT. This study presents a short-term forecasting framework tailored for dock-based shared micro-mobility services, exemplified through a case study for Capital bike-sharing in Washington D.C. Addressing the complex patterns and the demand variations caused by external uncertainties, we identify four key sources of forecasting inputs: seasonality, weather conditions, passenger flows from connected metro services, and previous demand observations. XGBoost and LSTMs are utilized as the basic models of our forecasting framework to capture the nonlinear dependencies between these features and demand. Additionally, we introduce a two-stage hurdle demand forecasting framework. It first estimates expected demand from contextual inputs, and then refine these predictions by accounting for short-term temporal demand fluctuation effects. Preliminary results show substantial forecasting improvements when these sources are integrated, with the highest accuracy achieved by hurdle models tailored to individual service points.

ABSTRACT. Increasing urban density and environmental concerns have driven the adoption of micromobility (MM) solutions like bicycles and e-scooters, which help mitigate issues like traffic congestion and pollution. These MM solutions are crucial for bridging the first and last-mile connectivity with public transport (PT) systems. Integrating MM with PT enhances urban mobility by using PT data to predict MM demand, optimize bike fleet management, and improve service reliability and user satisfaction. Recent studies utilize both traditional statistical methods and advanced deep learning techniques to predict MM demand using PT data. Traditional methods such as Poisson regression and SARIMA, though interpretable, struggle with complex urban dynamics. In contrast, advanced methods like Random Forests, Gradient Boosting Machines, and deep learning models such as Graph Convolutional Networks (GCNs) are better at handling non-linear relationships and spatial dependencies. The paper proposes a novel approach using a Multi-channel Spatio-Temporal Graph Convolutional Network (MC-STGCN) that integrates PT checkout data with MM demand, also considering external factors like weather, to enhance prediction accuracy and adapt to varying demand volumes. Notably, method performance measures, e.g., Mean Absolute Error (MAE) show improvement for our proposed methodology, which includes bike pickups, subway checkouts, and weather data (MM-PT-W), compared to the other configurations.

ABSTRACT. The advent of advanced traffic data collection technologies, such as loop detectors and GPS devices, has revolutionized traffic management by enabling the generation of vast data amounts. This has facilitated the development of sophisticated machine learning (ML) models for real-time traffic incident prediction. These models, which are trained on historical traffic data, play a pivotal role in forecasting traffic incidents. However, the inherent variability and unpredictability of traffic dynamics and incident patterns necessitate that these models not only predict incidents with high accuracy but also gauge the certainty of their predictions. In critical safety applications, uncertainty estimation is vital, enabling traffic management authorities to measure the confidence levels of predictions and make informed decisions. This capability is essential for effective decision-making in road safety and for implementing these models in production environments. A particularly relevant use case for uncertainty estimation arises during adverse and rare event conditions, such as severe weather, which significantly alters road conditions and driver behavior. For example, if a model predicts a high risk of accidents on a specific road link during a storm but with notable uncertainty, traffic managers might choose to issue a general caution instead of a full closure. This informed strategy optimizes safety without unnecessary traffic disruption by considering both the predicted risks and the confidence levels of those predictions. Research in traffic incident prediction increasingly utilizes Machine Learning (ML), especially Deep Learning techniques such as Graph Neural Networks (GNNs), due to their efficacy in handling complex traffic data patterns. GNNs are well-suited for modelling the graph structures of road networks, a strength demonstrated by Wang et al. (2021a)’s GSNet for regional risk assessment and further applied in city-wide risk evaluations (Wang et al., 2021b). Our previous work advances these models through the Multi-structured Graph Neural Network (MSGNN) (Tran et al., 2023), which integrates diverse data sources for ’network-wide’ predictions. However, the dynamic nature of traffic conditions and behaviors poses a challenge to the reliability of ML models trained on static datasets, potentially leading to obsolete or overly confident predictions in real-world critical applications. Thus, ensuring the reliability and timely applicability of these models is critical, mirroring trends in safety-critical fields like healthcare (Dolezal, 2022) and autonomous driving (Dong et al., 2023), where uncertainty quantification and explainable AI are key. Motivated by the necessity for reliable Traffic Incident Prediction (TIP), we diverge from existing works that emphasize model capability, focusing instead on reliability through our novel Uncertainty Aware Traffic Incident Prediction (UATIP) framework. This enriches predictive models with an Uncertainty Estimation (UE) ability, enhancing real-time interpretability and reliability—a facet not thoroughly addressed in current research. UE aids in discerning highly confident predictions from those warranting caution, thus bolstering the practical utility of traffic prediction models. Our experiments on MSGNN with varied real-world data highlight the UE’s role in augmenting prediction reliability. This advancement portends significant contributions to Explainable AI (Dong et al., 2023) in TIP and supports the application of Trustworthy Transfer Learning (TTL) (Shen et al., 2023), promising enhanced accuracy and robustness in a diverse array of traffic scenarios. Further details and results will be presented in the full paper.

ABSTRACT. Aircraft boarding is always a critical process and mostly driven by the behavior of the passengers. The duration of boarding is influenced by their air travel experience and their willingness or ability to follow boarding procedures. In fact, there is no feedback from the cabin about the current situation or indications of future conditions. It is reasonable to describe the cabin as a black box in today's digital age. This is more surprising given that most passengers carry their own transceivers (mobile devices) and thus technological solutions for feedback from the cabin do exist. From a research perspective, we want to answer three questions. (1) What sensor information needs to be provided at what quality? (2) How to cover domain knowledge in the ML models using complexity metrics? (3) How accurately could aircraft boarding times be predicted?

ABSTRACT. As any complex system, rail networks are vulnerable to delays and disruptions. Rescheduling actions are usually carried out neglecting the presence of uncertainties with a substantial underestimation of the risks. In this paper, given a solution for the real-time train rescheduling problem computed using deterministic methods, we asses the risk caused by the presence of uncertainty on the dwell time duration. We assume that the dwell time uncertainty to be represented by fuzzy numbers and we evaluate the Conditional Value-at-Risk (CVaR) of the makespan in a temporal network with fuzzy activity durations on the arcs. Computational tests carried out on realistic problems show how the proposed approaches are able to compute the exact fuzzy CVaR values of the makespan within extremely short computation times.

ABSTRACT. Investigation of train delay evolution and prediction is one of the hot topics in railway operations research. This is required for the effective and efficient realization as well as the quality improvement of train services. A wide range of papers investigating the train delay evolution and prediction problems has been published in the literature. The Markov chains is amongst the techniques used for this purpose. The interpretability, transparency and stability support its applicability in this respect. The workability of the discrete time Markov chains approach has been shown in the previously published papers. In this paper we extended this research towards the continuous time Markov processes to investigate the stochastic evolution of train delays using the continuous probability functions of time. To the best of our knowledge, this modelling approach applied for train delay investigation is used for the first time in the literature. It is shown in the numerical tests (despite limited in this extended abstract) that the promising results have been obtained in the analysis, evolution and prediction of train delays.

ABSTRACT. This paper proposes exact numerical solution methods for eco-driving problems of a single train. To find global optimal solutions of the problem, the main difficulty lies in the non-convex time dynamics. We propose a convex relaxation for the non-convex dynamics, and a rigorous mathematical proof of the equivalence between the reformulated model and original model could be provided. Although the speed-dependent force constraint of the reformulated model is bilinear, it can be efficiently solved to global optimum by off-the-shelf solvers. Numerical experiments are conducted to investigate performances of the proposed method in terms of solution quality and computational efficiency. Computational results show that the proposed method can efficiently deliver global optimal solutions and outperforms the benchmark method.

ABSTRACT. Power peaks are an undesirable phenomenon occurring in railway networks when multiple electric trains require a large amount of power simultaneously, putting pressure on the power grids. Furthermore, with increased traffic supply in recent years, the chances of having power peaks above certain levels increase, while at the same time, upgrading or over-dimensioning energy supply systems is very costly. One solution for this is fine-tuning of timetables to minimize power peaks. In this paper, we propose a novel simulation-based local search approach for the mitigation of power peaks in railway networks and rescheduling in real-time using two types of control measures: departure time shift and traction power limitation. The framework combines a detailed event-driven train traffic simulator with a local search strategy to efficiently explore and rank combinations of control measures, balancing the mitigation of power peaks and adherence to the timetables. We demonstrate the performance of the approach developed in a real-life case in a part of the Swiss Federal Railways network. The results show its potential to mitigate peaks and reduce energy costs with slight rescheduling actions in real-time with a very fast solving time, delivering nearly optimal solutions and negligible passenger inconvenience.