3D Object Detection for Road Safety at Urban Intersections
Date
2024
Authors
Howe, Matthew Robert
Editors
Advisors
Reid, Ian
Journal Title
Journal ISSN
Volume Title
Type:
Thesis
Citation
Statement of Responsibility
Conference Name
Abstract
Road safety is a critical issue, with road trauma resulting in 1.25 million fatalities and 50 million injuries annually worldwide. Using crash data and traditional methods, it can take years and even decades to collect enough data to evaluate the safety of complex road infrastructure, such as intersections. Long data collection times necessitate a proactive approach using surrogate measures of safety to address risks preemptively. This thesis delves into improving the accuracy of automatic measurements of surrogate safety indicators (SSI), adapting state-of-the-art computer vision techniques for surrogate safety analysis (SSA), primarily focused on intersection safety. I highlight potential error propagation in prior SSA methods in the literature review. My first significant contribution was to evaluate the error propagation of 2D and 3D object detection models for SSA. To achieve this, I built a bespoke conflict simulator to generate ground-truth conflicts that can be extracted by each object detector method. My research identified how minor inaccuracies can dramatically skew SSIs and emphasised the efficacy of 3D object detection in analysing conflicts at urban intersections. To address this issue, I introduce a technique for fine-tuning 3D object detectors using a novel wide baseline multi-view dataset. Traditional training methods for 3D object detectors, constrained by extensive data labelling and expensive sensor suites, are often inaccessible to road safety researchers. Mymethod usesweak supervision to fine-tune models trained on autonomous vehicle datasets for traffic observation cameras, achieving high-accuracy localisation. However, this work has the limitation of requiring multi-view data. To solve this, I make a further contribution by showing that other expected single-view cues can also be used for self-supervision to improve the performance of a 3D object detector. By integrating road user movement patterns, vehicle dynamics, and temporal consistencies, this approach fine-tunes 3D object detector performance utilising a single monocular camera view of an intersection. My work sets a promising trajectory for future road safety research and improves the accuracy, reliability, and accessibility of 3D object detection for SSA.
School/Discipline
School of Computer and Mathematical Sciences
Dissertation Note
Thesis (Ph.D.) -- University of Adelaide, School of Computer and Mathematical Sciences, 2024
Provenance
This electronic version is made publicly available by the University of Adelaide in accordance with its open access policy for student theses. Copyright in this thesis remains with the author. This thesis may incorporate third party material which has been used by the author pursuant to Fair Dealing exceptions. If you are the owner of any included third party copyright material you wish to be removed from this electronic version, please complete the take down form located at: http://www.adelaide.edu.au/legals