Time-Series-Based Personalized Lane-Changing Decision-Making Model.

Area of Science:

• Autonomous Driving Systems
• Human-Vehicle Interaction
• Machine Learning for Transportation

Background:

• Autonomous driving systems are shifting from human adaptation to vehicle adaptation.
• Improving the adaptability of autonomous driving systems to human drivers is crucial for widespread adoption.
• Personalized driving behavior models are needed to enhance human-vehicle interaction.

Purpose of the Study:

• To propose a time-series-based personalized lane change decision (LCD) model.
• To enhance the adaptability of autonomous driving systems to diverse human driving styles.
• To accurately predict lane change behavior based on individual driver characteristics.

Main Methods:

• Utilized Gaussian Mixture Model (GMM) for unsupervised clustering of driving styles based on speed, acceleration, and headway.
• Developed a gain function to quantify the interaction between the subject vehicle and surrounding traffic.
• Employed a Long Short-Term Memory (LSTM) recurrent neural network (RNN) for the personalized LCD model, incorporating driving style and interaction gain.

Main Results:

• Achieved high performance in lane change behavior prediction with an accuracy of 0.965, F1 score of 0.951, and macro-AUC of 0.983.
• Demonstrated significantly superior performance compared to other mainstream models.
• Successfully captured and replicated the lane change decision behaviors of different human drivers.

Conclusions:

• The proposed personalized LCD model effectively adapts autonomous driving systems to individual human drivers.
• The method offers a significant advancement in creating more intuitive and personalized autonomous driving experiences.
• This approach holds promise for improving the safety and acceptance of autonomous vehicles.