A CNN-LSTM Car-Following Model Considering Generalization Ability.

Area of Science:

• Traffic Engineering
• Artificial Intelligence
• Transportation Systems

Background:

• Understanding car-following dynamics is crucial for traffic flow optimization.
• Existing models struggle to capture the complex interactions and heterogeneity in real-world traffic.

Purpose of the Study:

• To develop a novel car-following model integrating Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks.
• To enhance the prediction accuracy and capture the heterogeneity of car-following behavior.

Main Methods:

• Extracted 400 car-following periods from natural driving and experimental databases.
• Developed a CNN-LSTM model where CNN analyzes vehicle dynamics and LSTM predicts speeds.
• Trained and tested the model against classical LSTM and Intelligent Driver Models.

Main Results:

• The CNN-LSTM model showed higher accuracy and better learning of traffic heterogeneity compared to other models.
• Accurately reproduced hysteresis phenomena in congested traffic.
• Demonstrated applicability to mixed traffic flow including adaptive cruise control vehicles.

Conclusions:

• The proposed CNN-LSTM model offers improved accuracy and generalization capabilities for car-following.
• It effectively models complex traffic dynamics, including congestion and mixed-vehicle environments.
• This approach holds promise for advanced driver-assistance systems and traffic management.