A deep clustering spatiotemporal framework for real-time urban traffic risk probability estimation.

Area of Science:

• Intelligent Transportation Systems
• Traffic Engineering
• Data Science

Background:

• Urban traffic dynamics present complex challenges for real-time risk assessment.
• Existing methods like Time to Collision (TTC) require extensive trajectory data, limiting scalability.
• Floating vehicle data offers a viable alternative for large-scale urban traffic analysis.

Purpose of the Study:

• To develop a real-time urban traffic risk probability assessment framework.
• To characterize traffic operational states using spatiotemporal grid representations and floating vehicle data.
• To enable proactive urban traffic risk prevention.

Main Methods:

• A Graph Attention Network Long Short Term Memory (GAT-LSTM)-based Spatiotemporal Autoencoder Neural Network (GL-SANN) was developed for latent risk feature extraction.
• A Deep Clustering Spatiotemporal Network (DCSN) with K-means clustering was used for traffic risk level assessment.
• A LightGBM model was employed for real-time risk identification, and a Spatiotemporal Graph Convolutional Risk Prediction (SGCRP) model for future risk inference.

Main Results:

• Traffic risk probability was classified into five levels, with higher risks during peak hours (50%-60% high-risk grids) compared to off-peak hours (approx. 37%).
• Intersections and high-traffic areas consistently showed elevated risk levels.
• The LightGBM risk identification model achieved 0.984 precision, and the prediction model achieved 0.974 precision with improved training speed.

Conclusions:

• The proposed framework effectively assesses and predicts urban traffic risk probability.
• Findings support proactive traffic risk management and the advancement of intelligent transportation systems.
• The methodology demonstrates the utility of spatiotemporal deep learning for real-time traffic safety analysis.