A spatio-temporal deep learning approach to simulating conflict risk propagation on freeways with trajectory data

Area of Science:

• Traffic Engineering and Safety
• Artificial Intelligence in Transportation

Background:

• Sudden vehicle maneuvers on freeways create propagating conflict risks, posing crash hazards and challenging advanced vehicle technologies.
• Simulating conflict risk propagation (CRPS) is complex due to intricate spatio-temporal dependencies and high-resolution data needs.

Purpose of the Study:

• To develop an effective method for simulating conflict risk propagation patterns on freeways.
• To introduce a novel Spatio-Temporal Transformer Network (STTN) for fine-grained conflict risk inference.

Main Methods:

• Introduced a conflict risk index aggregating surrogate safety measures (SSMs) like Modified Time-To-Collision (MTTC), Proportion of Stopping Distance (PSD), and Deceleration Rate to Avoid a Collision (DRAC).
• Proposed a Spatio-Temporal Transformer Network (STTN) with multi-head attention and spatial/temporal learning components to model spatio-temporal dependencies.
• Validated the STTN against benchmark models (LSTM, CNN, ConvLSTM) using real-world freeway datasets.

Main Results:

• The STTN demonstrated superior performance in simulating conflict risk propagation compared to traditional machine learning models.
• The PSD-based model showed robust performance across tasks, while the DRAC-based model effectively captured spatio-temporal heterogeneity.
• The attention mechanism within the STTN proved effective for CRPS tasks.

Conclusions:

• The proposed STTN offers a powerful tool for accurate conflict risk propagation simulation on freeways.
• Findings provide valuable insights for developing advanced traffic safety warning systems, freeway management, and driver assistance systems.