Recognizing autonomous driving disengagement scenarios using the transferable knowledge from human driver's EEG...

Area of Science:

• Neuroscience
• Robotics
• Human-Computer Interaction

Background:

• Autonomous driving (AD) systems struggle with rare, complex disengagement scenarios requiring human intervention.
• Human errors remain a significant factor in traffic accidents, highlighting the need for improved AD safety.

Purpose of the Study:

• To develop a transfer learning framework that integrates human electroencephalogram (EEG) cognitive data to enhance AD's recognition of disengagement scenarios.
• To improve the AD system's ability to understand and respond to critical situations where human takeover is necessary.

Main Methods:

• Collected EEG data from participants in both manual driving (MD) and AD supervision roles using a driving simulator.
• Developed a transfer learning framework incorporating a conditional maximum mean discrepancy (CMMD) function to align cognitive data between MD and AD domains.
• Trained a recognition model using the common feature space derived from EEG data.

Main Results:

• The proposed model achieved an 80% recognition rate for typical disengagement scenarios (e.g., static obstacles, intersection conflicts, vehicle cut-ins) using only 30% of AD training data.
• Transfer learning using EEG data improved recognition accuracy by 21.2% compared to models trained solely on AD data.
• The model effectively transferred knowledge from the cognitively demanding MD domain to the less demanding AD domain.

Conclusions:

• Integrating human EEG cognitive data via transfer learning significantly enhances AD's ability to recognize disengagement scenarios.
• Accurate recognition of disengagement types allows AD systems to implement appropriate safety mechanisms or takeover prompts, reducing accident risks.
• This approach offers a promising method for improving the safety and reliability of autonomous driving systems.