Accurate traffic accident detection is crucial to improving road safety conditions and route navigation, and to making informed decisions in urban planning among others. This paper proposes a Bayesian quickest change detection approach for accurate freeway accident detection in near–real–time based on speed sensor readings. Since post–accident conditions are hardly known, a maximum likelihood method is devised to track the relevant unknown parameters over time. Four aggregation schemes are designed to exploit the spatial correlation among sensors. Evaluation on real–world data collected from the I405 freeway in the Los Angeles County demonstrates significant gains as compared to the state–of–the–art in terms of average detection delay and probability of false alarm by up to 58.9% and 81.5%, respectively.