AI and Smart Mobility

Smart mobility is rich in data. There is enormous potential within the transport sector for using AI to address pressing challenges in sustainability and resilience, as well as fulfilling broader societal objectives of inclusion and equity. AI is already commonplace in many everyday Transport Systems and will become even more prevalent as technology evolves, particularly with rapid developments in Autonomous Vehicles. Our Fellows will discuss state of the art and propose how Autonomous Vehicles and future Transport Systems based on AI implementation will better meet and shape the growing and ever-changing demand for mobility while reducing its carbon footprint to meet net-zero goals.

Convened by: Andrew Morris, Haitao He

IAS Visiting Fellows in residence:

Mohamed Abdel-Aty, University of Central Florida
Kay Axhausen, ETH Zurich
Sebastien Glaser, CARRS-Q, Queensland University of Technology
Nigel Wilson, MIT CEE
Andry Rakotonirainy, CARRS-Q, Queensland University of Technology
Juan de Dios Ortúzar, Department of Transport Engineering and Logistics, Pontificia Universidad Catolica de Chile
Xuesong Wang, Tongji University

For more details on the IAS Annual Theme of AI: Facts, Fictions, Futures, please click here.

In order to attend this in-person event, please click here.

Programme:

Professor Mohamed Abdel-Aty, University of Central Florida

Next Generation Roads: AI, Big Data & Computer Vision Applications

This presentation will focus on the safety and operation benefits of AI, big data, technology, connected and Automated Vehicles (CAV) and active traffic management (ATM). The advent of AI and big data enables the real-time analysis of traffic safety and operation. By integrating multiple data sources, the data could help us understand the relationship between the presence of traffic conflicts and the real-time crash contributing factors (e.g. volume, speed, traffic control status, weather, etc), and quantify the impact of these factors on real-time crash risk. The same data could also be used to evaluate congestion and travel time reliability in real time. We also provide some insights from our studies that are related to ATM, using AI/big data analytics to improve safety and the potential of Connected and Automated Vehicles (CAV). For now, it is important to determine the expected effect that CAV technologies would have in reducing crashes. This is meaningful because we could provide important guidance for CAV-related policies, research, resource allocation, manufacturing and promotion of these systems.

This presentation would address the following topics:

• (Pro) Active Traffic Management (PTM)
• Machine Learning
• Big Data
• Real-Time Applications
• Computer Vision and Machine Learning
• Visualization

Professor Sebastien Glaser, CARRS-Q, Queensland University of Technology

AI and AV: understanding local context 

If the control task of an automated vehicle is dealt with physics and automatics, the mechanisms to understand the context and adapt the decision-making process are now driven by Machine Learning. However, the concept of Machine Learning demonstrates some limitations when it must manage local situations. This presentation identifies some challenges and shows preliminary results from the Cooperative and Highly Automated project (an iMOVE Australia, Queensland Department of Transport and Main Roads and Queensland University partnership).

Professor Xuesong Wang, Tongji University

Traffic Safety Analyses Using Shanghai Naturalistic Driving Study Data

Shanghai Naturalistic Driving Study (SH-NDS) was jointly conducted by Tongji University, General Motors (GM), and the Virginia Tech Transportation Institute (VTTI). Five GM light vehicles equipped with Strategic Highway Research Program 2 (SHRP2) NextGen data acquisition systems (DAS) were used to collect real-world driving data. SH-NDS data are applied for crash causation analysis, driving behavior analyses, and vehicle control modeling. Vehicle kinematic triggers were utilized to identify crash and near-crash events (CNCs). Pre-crash scenarios were identified using the Pre-Crash Scenario Typology. In-depth investigations of CNCs in the same scenario were analyzed to determine the causes of crashes using the systematic crash causation derivation framework. Car following model and lane changing model were calibrated for Chinese drivers for different roadway types (i.e., arterials, urban expressway, freeways). Using SH-NDS data, safe, efficient, and comfortable velocity control models were developed based on reinforcement learning for autonomous driving. Responsibility Sensitive Safety (RSS) models for car-following, lane changing, vehicle to bicycles, vehicle to pedestrian critical events were calibrated and evaluated.

Professor Kay Axhausen, ETH Zurich

Is cycling a way out of transport policies' dead end?

Bicycles and e-bikes, in particular, are competitors for the car, but also public transport in urban areas. The limited amount of road space now made available to them and the generally sorry maintenance of the generally limited amount of dedicated facilities make them unattractive for a large share of the population. The high historical usage of the cycle tells us, that there is no law of nature behind the currently low shares.
Transport policy has to guide the transport system to a net zero future by 2050, i.e. removing the about 30% share of CO2 due to the transport sector. It should also continue to increase (maintain) the accessibility levels fairly. All major policy levers have known downsides, which make them hard to impossible to implement: Urban road building is both expensive and invites unwanted modal shifts; all pricing alternatives (road use, congestion charging, parking, transit peak pricing) are too unpopular to implement unless special circumstances prevail; electric cars are helpful, but still produce too much CO2 in either production or use; automated cars are likely to produce so much additional travel, that the net balance might be negative, as more persons can drive in these cheap to use and comfortable vehicles. Small public transport vehicles (4-8 seaters) for on-demand and pooled services are generally uneconomical away from rare concentrated demands, such as a train arrival, low private car ownership, or a group of friends travelling together.
In this dead end it is necessary to envision positive futures. The IVT is currently testing a scenario where cycles and other slow modes are allocated 50% of the existing road space to make it a safe and comfortable option. We will check all of the technical constraints, especially due to transit, and assess the required behavioural changes. The e-bike-city is intended to be the focus of discussion to move the discussion towards a new starting point for the necessary changes.

Professor Nigel Wilson, MIT CEE

AI and Smart Urban Public Transport

On behalf of Jinhua Zhao, Haris Koutsopoulos, Shenhao Wang, Joseph Rodriguez, Qingyi Wang, and Dingyi Zhuang

In this talk we illustrate the potential of AI application in public transit using two examples: 1) Uncertainty Quantification of Spatiotemporal Transit Demand with Probabilistic Graph Neural Networks; 2) Cooperative Bus Holding and Stop-skipping: A Deep Reinforcement Learning Framework. The first study examines transit demand prediction. Recently deterministic neural networks have significantly improved the prediction accuracy of average travel demand but they have largely ignored the uncertainty that inevitably exists in transit demand prediction. Our study proposes a framework of probabilistic graph neural networks (GNN) to quantify the spatiotemporal uncertainty of travel demand. Applying it to the transit demand prediction in Chicago for pre- and post- COVID periods, we found that the probabilistic GNNs can successfully describe the spatial and temporal uncertainty patterns at a high spatial and temporal resolution. The second study focuses on the bus control problem that combines holding and stop-skipping strategies. We formulated it as a multi-agent reinforcement learning (MARL) problem with a design of the state and reward function that increases the observability of the impact of agents’ actions during training. An event-based mesoscopic simulation model is built to train the agents. We evaluate the proposed approach in a case study from the Chicago transit network. The results show that the proposed method not only improves the level of service but it is also more robust towards uncertainties in operations such as travel times and operator compliance with the recommended action. We conclude the talk with an initial research agenda on the AI applications in public transit.

Professor Andry Rakotonirainy, CARRS-Q, Queensland University of Technology

A human-centric eXplainable Automated Vehicle (XAV)

The presentation will address the inability of Automated Vehicles, powered by Artificial intelligence, to self-explain their behaviours. I will present a project which applies novel multidisciplinary methodologies in a real-world self-driving setting to formalise the essence of driving explanations. It explores the when, why and how a driver is seeking an explanation and what type of automated explanation is truly human-interpretable.

Professor Juan de Dios Ortúzar, Department of Transport Engineering and Logistics, Pontificia Universidad Catolica de Chile

AI and Smart Mobility - Contributions from the Deep South

This presentation succinctly reviews AI contributions to smart mobility in Chile. From the pioneering use of smart card data to estimate public transport matrices, to several on-going research projects related to AI use in modelling the perceptions of urban space, public transport satisfaction, the behaviour of public transport drivers, and the incorporation of psychophysiological indicators in the analysis of transport systems. We also mention some governmental projects and new proposals to use AI in travel demand and traffic management in the city, highlighting their expected benefits.