A deep transfer learning approach for Real-Time traffic conflict prediction with trajectory data

Area of Science:

• Transportation Engineering
• Artificial Intelligence
• Traffic Safety

Background:

• Real-time traffic conflict prediction is crucial for proactive traffic safety systems.
• Existing models lack transferability, showing performance degradation when applied to new locations due to varying traffic environments.
• This limitation hinders the widespread adoption of predictive traffic safety solutions.

Purpose of the Study:

• To propose a novel deep transfer learning approach to enhance the transferability of real-time traffic conflict prediction models.
• To evaluate the effectiveness of different deep learning architectures (Gated-Transformer, FCN, LSTM-FCN, MLSTM-FCN) for feature extraction in traffic conflict prediction.
• To assess the performance improvement gained by applying transfer learning to models trained on different traffic domains.

Main Methods:

• Developed a real-time conflict prediction framework using trajectory data and considering temporal traffic flow variations.
• Employed Gated-Transformer, FCN, LSTM-FCN, and MLSTM-FCN as backbone networks for feature extraction.
• Implemented an independent transfer learning architecture based on maximum mean discrepancy to assess domain similarity.
• Empirically evaluated models using the exiD dataset, differentiating source and target domains.

Main Results:

• The Gated-Transformer model demonstrated superior performance in feature extraction and prediction compared to baseline models, achieving an F1 score of 0.864 and AUC of 0.980.
• The transfer learning architecture significantly improved the predictive performance of models applied to new domains.
• For the Gated-Transformer model, transfer learning led to an 11.9% increase in F1 score and a 10.2% increase in AUC.
• Sensitivity analysis recommended optimal parameters: a sliding time window of 6 seconds and a prewarning time of 5 seconds.

Conclusions:

• Deep transfer learning offers a reliable and effective method for improving the transferability of real-time traffic conflict prediction models.
• The Gated-Transformer architecture shows strong potential for traffic conflict prediction and domain adaptation.
• Findings provide valuable insights for developing practical traffic safety warning systems, especially in vehicle-to-infrastructure environments.