Customizing Trustworthy Machine Learning and Advanced Computing Methods for Cooperative and Equitable Traffic Systems

Abstract

Urbanization offers opportunities for better lives and amenities globally, yet it brings considerable challenges. The combination of the network congestions, the inequity complaints, the unbalanced supply and demand of traffic resources, and gaps in integrating traffic needs with emerging technologies are converging to expand our understanding of smart traffic systems from a holistic point of view. To address these challenges, it is necessary to develop smart, cooperative, and equitable next-generation transportation systems. These systems need to integrate the machine intelligence to accomplish three key tasks: capturing real-time, accurate traffic information through sensors, integrating advanced data processing and analytical tools, and providing trustworthy services to roadway users without discrimination. This dissertation explores the customization of vital technologies and requisite methods for constructing trustworthy next-generation traffic systems. Specifically, the thesis advocates for a tripartite framework that aligns trustworthiness across three essential components: the data, methodological, and system levels, which denotes the intelligent transportation systems that are more cooperative, accurate, reliable, and less discriminatory. These systems and methods exploit the combination of tailored machine learning methods, along with distributed computing systems, to propose promising solutions for three key tasks: 1) Cooperative Traffic Perception – develop a privacy-aware cooperative sensing system through learning cross-sensor visual representations. A pioneering edge-empowered cooperative multi-camera perception system has been proposed and tested based on Internet of Things architecture. 2) Multimodality Data Fusion – pertain to advancing the customization of machine learning and feature fusion methodologies for multimodal data representations, with the aim of providing accurate and timely future traffic information for both roadway users and managers. Currently, the traffic forecasting approaches face obstacles such as limitations supporting multiple sensor modalities, insufficient traffic domain knowledge integrations, and poor generalization across heterogeneous traffic network structures. The works develop customized representation learning methodologies to predict traffic patterns of different scales and input formats. These proposed methods have outperformed sixteen cutting-edge machine learning approaches in predicting route speeds for hundreds of users, parking lot occupancy for nineteen lots, and network-scale congestion in cities worldwide. 3) Demonstrating Pilot Cooperative and Equitable Traffic Infrastructure Systems for mobility, safety, and equity enhancement. The systems utilize customized representation learning and edge computing approaches to address critical transportation needs. These include developing a human-machine cooperative automatic traffic signal assistance system for Vulnerable Road Users (VRU) and creating a vehicle recognition framework that takes unbalanced data samples into consideration. The outstanding performance of these representation learning methods significantly enhances both inter and intra-cooperation within traffic systems, from the perspectives of sensors, data, system, and their methodologies. Furthermore, advancements in the precision, generalization and privacy preservation underscore the substantial benefits of integrating transportation domain needs into machine learning and advanced computing research, which is pivotal for the future development of trustworthy smart traffic systems.