Cockpit-Llama: Driver Intent Prediction in Intelligent Cockpit via Large Language Model.

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Automotive Engineering

Background:

• The automotive cockpit is transitioning from reactive to proactive interaction, necessitating accurate driver intent prediction.
• Enhancing proactive interaction requires understanding driver behavior and cognitive states within the cockpit environment.

Purpose of the Study:

• To introduce Cockpit-Llama, a novel language model for predicting driver behavior intent.
• To improve the accuracy and rationality of driver intent predictions by leveraging multi-attribute cockpit data.

Main Methods:

• Developed Cockpit-Llama, a language model for driver intent prediction based on actions, historical interactions, and environmental states.
• Constructed a multi-attribute cockpit dataset including driver emotional, driving, vehicle, body, and environmental states.
• Fine-tuned the Llama3-8b-Instruct model using the Low-Rank Adaptation (LoRA) method for efficient parameter optimization.

Main Results:

• Cockpit-Llama demonstrated superior prediction performance compared to advanced methods on the multi-attribute cockpit dataset.
• Achieved high scores: BLEU-4 (71.32), ROUGE-1 (80.01), ROUGE-2 (76.89), and ROUGE-L (81.42).
• Exhibited significant relative improvements over ChatGPT-4, including 183.61% for ROUGE-1 and 201.27% for ROUGE-L.

Conclusions:

• Cockpit-Llama effectively predicts driver intent, enhancing proactive interaction in intelligent cockpits.
• The model significantly improves the reasoning and interpretative capabilities of automotive systems.
• The developed multi-attribute dataset and LoRA fine-tuning contribute to efficient and accurate driver behavior modeling.