Utilizing the eigenvectors of freeway loop data spatiotemporal schematic for real time crash prediction

• The shortcomings of the previous study on accident precursor models include multicollinearity, failing to reflect the overall traffic flow status before the crash, high quality loop data input requirement, and poor portability.
• This paper proposes a new method to model the real time crash likelihood based on loop data through schematic eigenvectors. Eigenvectors and eigenvalues of the spatiotemporal schematics were extracted to represent traffic volume, occupancy and speed situation before the crash occurrence. By setting the vectors in crash time as case and those at crash free time as control, a logistic model is constructed to identify the crash precursors.
• Results show that both the eigenvectors and eigenvalues can significantly impact the accident likelihood. The eigenvectors of speed schematics have the most significant parameter estimates. The third eigenvector in the 6 × 6 speed matrix mostly impacts the crash likelihood and the speed variation in the section near the crash site is one of the major risks to the crash.
• Compared to the previous study, the proposed model has the advantage of avoiding multicollinearity, better reflection of the overall traffic flow status before the crash, less loop data input requirement, and improving portability.

The concept of crash precursor identification is gaining more practicality due to the recent advancements in Advanced Transportation Management and Information Systems. Investigating the shortcomings of the existing models, this paper proposes a new method to model the real time crash likelihood based on loop data through schematic eigenvectors. Firstly, traffic volume, occupancy and density spatiotemporal schematics in certain duration before an accident occurrence were constructed to describe the traffic flow status. Secondly, eigenvectors and eigenvalues of the spatiotemporal schematics were extracted to represent traffic volume, occupancy and density situation before the crash occurrence. Thirdly, by setting the vectors in crash time as case and those at crash free time as control, a logistic model is constructed to identify the crash precursors. Results show that both the eigenvectors and eigenvalues can significantly impact the accident likelihood compared to the previous study, the proposed model has the advantage of avoiding multicollinearity, better reflection of the overall traffic flow status before the crash, and improving missing data problem of loop detectors.