4. Estimating Driving Performance Based on EEG/ICA Power Spectrum
4.4. Drowsiness Measurement
4.6.7. AFSM-based Driving Performance Estimation/Prediction
We also proposed a novel adaptive feature selection mechanism (AFSM) to solve the reliable and sorting problem of ICA components based on the correlation analysis between the time-frequency power spectra of ICA components and the driving performance. To reduce the
components to be two and five, respectively, by using AFSM. Table 4-9 shows the number of ICA components and optimal frequency bands selected manually and those by AFSM. Note that both selected ICA components and the frequency bands are almost the same but slightly different for each subject between manually selecting and AFSM. To verify the correctness and effectiveness of the AFSM method, the selected log bandpower spectra of the ICA components in these critical bands were feed as the input features of the linear regression models. We also used the Self-cOnstructing Neuro-Fuzzy Inference Network (SONFIN) [94]
model to estimate and predict the individual driver’s driving performance by taking the advantages of fuzzy reasoning and learning abilities, and flexibility of neural networks. Fig.
4-21 shows the driving performance estimation for training/testing sessions of subject 3, based on SOFNIN models (red line) with input features selected by AFSM method according to Table 4-10, overplotted against actual driving performance time series for the session (blue line). The correlation coefficient between the two time series is r=0.96 in the training session and r=0.94 in the testing session.
Table 4-10 shows the comparison results of driving performance estimation. Although the correlation coefficients between the two time series based on AFSM methods using linear regression models are somewhat lower than those selected manually. The adaptive feature selection mechanism has the advantages of saving time, and cost when the whole system is applied for on-line alertness monitoring. Table 4-11 shows the estimating results based on AFSM methods using SONFIN. Compared to the results using linear regression models, using fuzzy neural network models can achieve higher estimating results as shown in Fig. 4-21 for subject-3, and can compensate the slightly loss using AFSM in real-time applications.
Table 4-9.
The optimal 2 ICA components and frequency bands selected manually and by AFSM corresponding to different subjects according to the higher correlation coefficients between log bandpower spectra and the driving performance.Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 ICA Components 17, 28 17, 8 11, 13 4, 5 22, 25 Manual
Freq. Bands 5-9 Hz 8-12 Hz 10-14 Hz 4-8 Hz 8-12 Hz ICA Components 17, 28 17, 8 11, 13 4, 5 22, 25 AFSM
Freq. Bands 4-8 Hz 8-12 Hz 10-14 Hz 5-9 Hz 9-13 Hz
Table 4-10.
Driving performance estimation using total 10 frequency bands in 2 dominant ICA components selected manually and by AFSM methods shown in Table 4-9, as input features of the linear regression models for five subjects.Sub-1 Sub-2 Sub-3 Sub-4 Sub-5 Average
Training 91% 91% 93% 89% 90% 90.8%
Manual Linear
Regression Testing 77% 89% 92% 86% 80% 84.8%
Training 88% 91% 93% 88% 84% 88.8%
AFSM Linear
Regression Testing 72% 89% 82% 80% 76% 81.8%
Table 4-11.
Driving performance estimation using total 10 frequency bands in 2 dominant ICA components selected by AFSM methods shown in Table 4-9, as input features of the linear regression models and SONFIN for five subjects.Sub-1 Sub-2 Sub-3 Sub-4 Sub-5 Average
Training 88% 91% 93% 88% 84% 88.8%
Linear
Regression Testing 72% 89% 82% 80% 76% 81.8%
Training 89% 92% 96% 87% 91% 91%
AFSM
(a) Training Result
(b) Testing Result
Figure 4-21. Driving performance estimation for training/testing sessions of subject 3, based on SOFNIN models (red line) with input features selected by AFSM method according to Table 4-9, overplotted against actual driving performance time series for the session (blue line). The correlation coefficient between the two time series is r=0.96 in the training session and r=0.94 in the testing session.
4.7. Conclusion Remarks
In this chapter, we demonstrated a close relationship between minute-scale changes in driving performance and the EEG/ICA power spectrum. This relationship appears stable within individuals across sessions, but is somewhat variable between subjects. Four computational approaches were proposed to select effective features for drowsiness estimation based on the compromise of computational cost and estimating accuracies. The first approach combined EEG power spectrum estimation, correlation analysis, PCA, and linear regression to continuously indirectly estimate/predict fluctuations in human alertness level indexed by driving performance measurement, deviation between the center of the vehicle and the center of the cruising lane. Our results demonstrated that it is feasible to accurately estimate driving errors based on multi-channel EEG power spectrum estimation and principal component analysis algorithm. The computational methods we employed in this study were well within the capabilities of modern real-time embedded digital signal processing hardware to perform in real time using one or more channels of EEG data. Once an estimator has been developed for each driver, based on limited pilot testing, the method uses only spontaneous EEG signals from the individual, and does not require further collection or analysis of operator performance. The proposed methods thus might be used to construct and test a portable embedded system for a real-time alertness monitoring system. The other two approaches used ICA, power spectrum analysis, correlation analysis, and the linear regression model in a virtual-reality based driving environment. Experimental results show that the proposed analysis methods are feasible to accurately estimate individual driving error accompanying loss of alertness by linear regression model with 10 subband log power spectra
electrodes of the corresponding ICA components. Average accuracies of training and testing session for 5 subjects are 88.2% and 79%, respectively. Although the accuracy is somewhat lower than those using ICA components, its does not require to collect more EEG channels data in testing session. Thus, this approach suggests a compromise between computational cost and estimation accuracy. Therefore, the proposed methods can be used to construct and test on an online portable embedded system for a real-time alertness monitoring system. In the last approach, we proposed a novel adaptive feature selection mechanism to solve the sorting problem of the ICA components and to extract useful frequency bands as input features.
Experimental results show that the average accuracies of training and testing session for five subjects can achieve high to 88.8% and 81.8% as well as 91% and 87%, by using linear regression model and fuzzy neural network models, respectively. Although the accuracy using AFSM-based linear regression model is lower than those selected manually, the computational methods we employed in this study were well within the capabilities of modern real-time embedded digital signal processing hardware to perform in real time alertness monitoring system.