Proactive Collision Risk Quantification in Multi-directional Traffic Interactions

Abstract

Road traffic crashes cause over a million deaths and tens of millions of injuries annually, with the majority occurring in complex multi-directional urban traffic interactions such as merging, turning, and crossing, rather than on high-speed motorways. These collisions rarely stem from a single error, but emerge from escalating conflicts, leaving a time window in which proactive intervention is possible. This thesis systematically develops a data-driven methodology to quantify collision risk in multi-directional urban traffic interactions, in a way that is context-aware, generalisable across scenarios, and scalable without relying on crash labels.

The research progresses from foundational measurement to large-scale risk modelling. First, a two-dimensional coordinate transformation is introduced to normalise longitudinal and lateral spacing between road users. This enables consistent microscopic measurement of interactions and macroscopic analysis of required road space via an interaction Fundamental Diagram (iFD). Building on this representation, a unified probabilistic framework for conflict detection is formulated. It conditions collision risk on interaction context, including motion kinematics and environmental factors. A statistical learning pipeline is then proposed to estimate continuous risk scores that generalise across scenarios and capture a long-tailed spectrum from mild conflicts to near-crashes. To scale up without annotated crash or near-crash events, the Generalised Surrogate Safety Measure (GSSM) is developed as a self-supervised approach that learns collision risk from abundant naturalistic driving data. Further, contrastive learning is explored to more effectively exploit fine-grained interaction patterns.

Experiments on real-world datasets show that lateral interactions utilise road space more efficiently than longitudinal ones, and that collision risk forms a continuum without a universal boundary between safe and unsafe interactions. The proposed context-aware methods achieve state-of-the-art risk detection accuracy and alert timeliness. Environmental factors such as rain, lighting, and surface conditions are shown to significantly impact collision risk. With increasing data in training and factors in consideration, extreme conflicts can be inferred more effectively from everyday interactions.

The proposed methods enable consistent measurement of road user interactions, adaptive conflict detection, unified collision risk scoring, and scalable learning in multi-directional traffic. In practice, the results can support applications in traffic management, advanced driving assistance and automated vehicles, real-time risk monitoring, and accelerated road safety policymaking. All these contribute to a broader shift from reactive to proactive road safety, aligning with the vision of eliminating traffic fatalities and creating more resilient urban transportation systems.