Date of Award
May 2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Civil Engineering
First Advisor
Sherif Gaweesh
Abstract
Traditional traffic safety analysis relies on crash data, which is inherently reactive and limited by underreporting and delayed insights. At the same time, existing trajectory extraction methods depend on LiDAR, radar, or calibrated camera systems, making them costly and difficult to scale. This study proposes a vision-based, calibration-free framework that leverages monocular CCTV footage for both trajectory estimation and proactive conflict analysis. A Bird’s Eye View (BEV) transformation combined with YOLOv8 Oriented Bounding Box (OBB) detection is used to extract accurate vehicle trajectories. A novel space-oriented bounding box formulation is introduced to better capture vehicle footprint under perspective distortion. The detection model achieves high performance with precision and recall exceeding 0.95 and [email protected]–0.95 of 0.92. Trajectory refinement through temporal smoothing and orientation correction produces stable and physically consistent motion profiles, enabling reliable estimation of speed and acceleration. Building on the extracted trajectories, conflicts are identified using Post Encroachment Time (PET) and Time to Collision (TTC), and severity is classified using Peak-over-Threshold Extreme Value Theory (EVT). The framework achieves 99% accuracy in conflict type detection when validated against manually annotated events. Results show that interactions with PET and TTC values greater than 1 second are consistently associated with low severity, while shorter time gaps correspond to higher-risk conditions. PET demonstrates greater robustness in CCTV-based applications, whereas TTC is more sensitive to trajectory estimation errors. The proposed framework provides a scalable, plug-and-play solution for proactive traffic safety monitoring using existing camera infrastructure. By integrating trajectory extraction, conflict detection, and severity modeling into a unified pipeline, it enables continuous identification of high-risk interactions and supports data-driven safety interventions aligned with modern traffic safety objectives.
Recommended Citation
Roy, Swaranjit, "Proactive Traffic Safety Assessment At High-Risk Intersections Using Computer Vision" (2026). Theses and Dissertations. 13976.
https://commons.und.edu/theses/13976