Thermal imaging and artificial intelligence for preventive road safety: a digital traffic shadow approach to conflict detection

Road safety planning has predominantly relied on retrospective accident data, a reactive approach that is statistically limited and fails to adequately capture latent risks present in everyday traffic. Proactive safety monitoring using video analysis is constrained by the weaknesses of RGB image data, including reduced robustness under certain weather and lighting conditions, as well as by significant data privacy concerns. Furthermore, the surrogate safety measures commonly employed often lack the level of interpretability required for engineering decision-making. This dissertation presents a data privacy-compliant system approach for preventive road safety analysis based on thermal imaging technology. The dissertation culminates in the development of a Digital Traffic Shadow, a dynamic 3D virtual representation that transforms raw thermal video data into operationally relevant and interpretable safety insights. The work addresses three methodological challenges through a cumulative series of studies covering machine perception, 3D reconstruction, and semantic conflict interpretation. First, a weakly supervised, incremental learning methodology is developed to overcome the lack of annotated thermal data and to enable robust, scalable object detection under heterogeneous camera perspectives. Second, the dissertation proposes a method for monocular 3D reconstruction. By integrating an efficient neural network with a geometric ray–plane intersection model, the system delivers precise 3D positions and orientations of road users in real time. Third, the extracted trajectories are embedded into the Knowledge PET3D system. This system extends classical surrogate safety measures by enriching kinematic indicators with semantic context. Through 3D interaction analysis, rule-based assessment of right-of-way relationships, and the use of a Transformer-based model for anomaly detection, the system provides interpretable evaluations of near-misses supported by linked video evidence and suitable for direct use in engineering practice. Overall, the results demonstrate that thermal imaging-based machine perception, 3D reconstruction, and knowledge-based interpretation can be integrated into a robust process chain. This enables continuous, data privacy-compliant monitoring under diverse environmental conditions. The developed Digital Traffic Shadow provides a foundation for automated, explainable, and scalable road safety audits, contributing to the development of intelligent infrastructure systems that support proactive traffic safety management.