Transformer-Conformer Ensemble for Crash Prediction Using Connected Vehicle Trajectory Data

Abstract

Crash prediction is one of the important elements of real time traffic management strategies. Previous studies have demonstrated the use of infrastructure-based detector data and UAV video to predict a crash in the near future. The main limitation of such data is limited coverage. In this work, we have used connected vehicle trajectory data that can have wide coverage as well as provide insight into the trajectory that might lead to a crash. The trajectory data was provided by Wejo which collects data from the manufacturer and was spaced at 3 seconds. GPS locations and their associated time series features such as speed, acceleration and yaw rate were used to feed into an ensembled Transformer and Conformer model. A voting classifier was used to obtain the output of the final model which achieved a recall of 76% and the false alarm rate of 30%. This study showed how connected vehicle trajectory data can aid in getting insight into crashes. While most previous studies focus on using aggregated data to estimate crashes, the proposed work shows that trajectory data mining can also provide competitive results.