Real-time recursive risk assessment for autonomous vehicles

Chia, Wei Ming (Dan) (2025) Real-time recursive risk assessment for autonomous vehicles. PhD thesis, University of Glasgow.

Abstract

The existing risk assessment coverage for Autonomous Vehicle (AV) deployment is insufficient for AV operations. The existing risk assessment is based on static processes, such as Hazard Analysis and Risk Assessment (HARA), which are performed during AV development. The hazard identification is based on prior lessons learned and know-how. The current risk assessment primarily focuses on vehicular malfunctions and does not assess AV's safety actions when hazards are detected during real-time operations. The static risk assessment is unable to measure AV control actions when a hazardous event is detected during AV operations. These planned control actions can also pose risks during AV operations without real-world testing.

This thesis proposes a real-time risk assessment for AV operations, comprising a Real-time Risk Assessment Framework (ReRAF) within the AV and a Real-time Risk Assessment for Cooperative mode (ReRAC) within the infrastructure. The realtime risk assessment operates in a continuous and recursive loop for safety improvement.

In ReRAF, a novel Dynamic Acquired Risk Assessment (DARA) algorithm is designed, developed and verified to provide a Predicted Risk Number (PRN) for the AV to assess the risk of its ability to mitigate hazards through its control actions. The PRN is derived using the scenarios' risk tag figures and the AV control tag figures. The risk tag figures are achieved by object detection, scene segmentation and probabilistic modelling, while control tag figures are derived from the AV’s parametric controls. The resulting PRN is a quantitative outcome of ReRAF in an objective end-to-end approach, without any human intervention within the AV. The DARA algorithm was tested in real-world AV operations with unregulated traffic scenarios that consist of vehicles and pedestrians. The accumulative PRN results over time can be used to identify potential hotspots and improve AV’s path planning. The DARA algorithm demonstrated the ability to use a single PRN to represent the real-time risk assessment of an AV by measuring the risk mitigation of the AV control parameters based on the detected risk from its camera.

For ReRAC, a novel Spatial-Temporal Risk Estimation Ensemble Technique (STREET) algorithm is designed, developed and verified to provide remote advanced risk warnings to the AV. STREET compute the environment’s risk tag figure and provides hazard identifications and warnings from the infrastructure viewpoint to the AV. Risk tag figures are obtained by first performing a risk zoning of the environment, followed by probabilistic modelling to convert the scene into a risk matrix. Object detection is then used to map the detected object onto the risk matrix to provide risk tag figures for the scene. ReRAC can also derive the time to collide, while hazard identifications and warnings are obtained by detecting pedestrians and/or vehicles in proximity within an intersection or road section of the scene. STREET provides an objective end-to-end approach without human intervention and was tested in unregulated traffic scenarios that provide advanced AV warnings using cooperative mode. STREET results demonstrated the ability to perform real-time conversion from qualitative image to quantitative risk tag figures from the infrastructure’s environment scene and time to collision to act as AV’s preemptive warning purposes. In addition, the STREET algorithm also illustrated the ability to detect hazard identifications and warnings with roads and intersections, such as vehicle-to-vehicle and vehicle-to-pedestrian hazards, as well as pedestrian warnings and vehicle warnings for intersections. These hazard identifications and warnings are potential risks when the AV arrives at the infrastructure location. They can also detect potential accidents and road congestions if the detection persists.

The combination of ReRAC and ReRAF provides complete coverage of safety enhancements for AV operations in real time, within and beyond the AV. With ReRAC operating as remote advanced warnings, the AV's resultant actions can be safer as it moves towards the ReRAC location. In parallel, the ReRAF continuously monitors and assesses the AV's real-time risk assessment and acts as a trigger if further improvements for safer actions are required for subsequent routes.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Supervisor's Name: Goh, Professor Cindy and Keoh, Dr. Sye Loong
• Date of Award: 2025
• Depositing User: Theses Team
• Unique ID: glathesis:2025-84978
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 24 Mar 2025 14:19
• Last Modified: 25 Mar 2025 09:02
• Thesis DOI: 10.5525/gla.thesis.84978
• URI: https://theses.gla.ac.uk/id/eprint/84978

Related URLs

• Article DOI
• Enlighten Publications Record
• Article DOI