Detection of Driver Braking Intention Using EEG Signals During Simulated Driving

Figures

• 

  Figure 1. Electroencephalogram (EEG) headset: (a) schematic of the 8-channel EEG circuit with dry active electrodes, (b) photographs of complete 8-channel EEG headset with 3-D printed flexible cap and the experimental setup for the virtual driving environment.

• 

  Figure 2. (a) Flowchart of the EEG-based braking-intention detection system, (b) flowchart of the labeling process for EEG data using the inertial measurement unit (IMU) sensor.

• 

  Figure 3. (a) Accelerometer and gyroscope data for 10 min of driving, with indicators for normal driving and emergency braking events; (b) features of the standard deviation of the windowed data: gyroscope features (top) and accelerometer features (bottom); (c) distribution of the time interval between two consecutive braking events generated by the driving simulator; (d) reaction time measured by the driving simulator.

• 

  Figure 4. Neural-network structure for the two-class classifier (normal driving vs. braking intention).

• 

  Figure 5. Comparison of the relative EEG power of eight channels between two situations: normal driving and braking intention: (a) delta; (b) theta; (c) alpha; (d) beta; and (e) gamma bands.

• 

  Figure 6. Performance comparison of the model between band power-based and autoregressive (AR)-based features for subject 4: (a) receiver operating characteristic (ROC) curves for the band power-based and AR-based classifiers; (b) system accuracy with respect to the number of hidden neurons in the artificial neural network (ANN) model; (c) cross-entropy loss during training and testing of the AR model-based classifier.

• 

  Figure 7. System accuracy with respect to the AR order number: (a) ROC curves for different AR classifiers with different AR order numbers; (b) average system accuracies during five runs of the neural-network model.

• 

  Figure 8. System accuracy with respect to the analysis step size for 10 subjects.

• 

  Figure 9. Labeling the braking intention samples within different pre-braking periods: (a) ROC curves; (b) AUC of the AR feature-based classifiers with respect to pre-braking period for 10 subjects.

• 

  Figure 10. Accuracy of the proposed system in online testing for 10 subjects.