A Bayesian multi-source data fusion approach to mitigate spatiotemporal uncertainties in traffic accident records

Abstract

Reliable and precise accident data are a crucial prerequisite for mitigating accident-related costs and improving the safety of the road traffic network. However, in practice, reported accident data often suffer from temporal and spatial uncertainties, creating a gap between when and where accidents are recorded and when and where their effects are observable in traffic flow and congestion. To address this challenge, we propose a Bayesian multi-source data fusion framework that reconstructs latent accident attributes by integrating heterogeneous evidence while explicitly quantifying uncertainty. Key contributions of this work are: (i) a latent-variable formulation under Bayesian inference that frames accident records as uncertain observations rather than deterministic inputs, (ii) a scalable multi-source data fusion approach that integrates matrix signaling information data and traffic measurements as complementary probabilistic evidence, and (iii) the resulting uncertaintyaware accident dataset can be directly used in data-driven models, such as Bayesian networks, leading to more reliable robust inference in safety and accident management applications.