Deep hybrid learning framework for spatiotemporal crash prediction using big traffic data

Abstract

The rapid growth in data collection, storage, and transformation technologies offered new approaches that can be effectively utilized to improve traffic crash prediction. Considering the probability of traffic crash occurrence vary due to the spatiotemporal heterogeneity, this study proposes a state-of-the-art deep learning-based model that incorporates spatiotemporal information for the short-term crash prediction, named as Deep Spatiotemporal Hybrid Network (DSHN). The model integrates Convolutional Neural Network (CNN), Long Short-term Memory (LSTM), and Artificial Neural Network (ANN) to incorporate the synergistic power of individual models. The study utilizes different data sources such as big traffic data collected from Paris road network sensors, weather conditions, infrastructure, holidays, and crash data. The results indicated that the proposed DSHN model outperforms the baseline models with an Area Under Curve (AUC) of about 0.800, an accuracy of 0.757, and a false alarm rate of 0.217. In addition, the importance of each data type is evaluated to investigate their impacts on the prediction performance of models. The sensitivity analysis results indicate that the road sensor data that includes average speed, vehicle kilometer traveled (VKT), and weighted average occupancy has the highest impact on the prediction accuracy.