Knowledge PET3D: An interpretable framework for 3D near-miss detection in thermal traffic video

Expert Systems with Applications, vol. 304, pp. 130821

Abstract

Traffic safety engineers often rely on retrospective crash data, limiting their ability to proactively identify systemic risks in road environments. To address this gap, this work presents Knowledge PET3D, a novel, privacy-preserving framework that enables automatic long-term traffic observation, detects safety-relevant interactions, and delivers interpretable video snippets to support informed engineering decisions. The system integrates monocular 3D detection, maneuver-specific rule modeling, and transformer-based anomaly detection to identify near-miss events from thermal video. The system filters non-informative interactions based on rule compliance and behavioral response, enabling interpretable conflict sets suitable for manual review. Compared to conventional PET2D and PET3D baselines, Knowledge PET3D achieves over 4x more true positives and reduces false positives by up to 93%. It delivers high precision across varied urban contexts (22% at a complex signalized intersection and up to 75% at a simpler yield-controlled site), while keeping conflict volumes verifiable by humans. The framework further achieves 87.0% correct classified maneuvers, 92.9% clustering accuracy, and 96.5% correct interpreted traffic rules. Knowledge PET3D advances traffic safety diagnostics by uncovering latent risks while maintaining engineering interpretability and operational scalability.

Authors 2

1. Arnd Pettirsch Aachen Institute of Highway Engineering

   RWTH Aachen University

   Affiliation as printed

   Institute of Highway Engineering, RWTH Aachen University, 52074 Aachen, Germany

2. Álvaro García corresponding Aachen Institute of Highway Engineering

   RWTH Aachen University

   Affiliation as printed

   Institute of Highway Engineering, RWTH Aachen University, 52074 Aachen, Germany