1.1. The importance of drowsiness detection
In previous study, the fatigue which caused drivers inattention or drowsiness, was the major risk factor for serious injury and death in car accidents [1-4]
National Sleep Foundation (NSF) reported that 60% of drivers had felt drowsy during driving, and 37% of the drivers had actually fallen asleep. The National Highway Traffic Safety Administration (NHTSA) also reported that at least 100,000 police-reported crashes were directly caused by drowsy driving in 2006 and leaded to 1,500 deaths, 71,000 injuries and $12.5 billion in monetary losses (National Sleep Foundation 2007 State of the States Report on Drowsy Driving).
Therefore, to early detect the drivers’ drowsiness and to help to keep the drivers’
alertness for avoiding the car accidents that caused by drowsiness are important to protect living safeties of people.
Drowsiness detection has been widely researched by varied measurements [5, 6] including the monitoring subject’s behavior and image based techniques and physiological signal-based system. The following sections would explain the advantage and limitation of these methods.
1.2. The Drowsiness detection index
1.2.1. The behavioral monitoring
Previous studies had shown that driver’s response performance is negatively relative to the drowsiness. The response performances were defined in terms of response time [7, 8], driving trajectories [9, 10] and patterns of drivers’
moving handle wheel [11, 12]. The limitation of behavioral monitory system is highly depended on driving behavior, experiences, road conditions, and all other environmental variables. But, previous have showed that behavioral performance
2
is opposite correlated with the driver’s alertness. Specifically, the subject’s response performances, which index by response time, are decreased along with the increases of drivers’ drowsiness [13, 14].
1.2.2. The image-based technique
The image-based technique detect the eye gaze position, eye closure or the head position by the video camera[15] to calculate the duration of eye gaze fixation and the eye closure or frequency of eye movement, eye blinking [16-18]
or head movement [19] for correlating the subject’s drowsiness level. However, the quality of recorded image is easily influenced by the environment [20], with which is necessary for the camera needed to interact.
1.2.3. The physiological signal based system
Several studies used the physiological signals, including the electrocardiograph (ECG), electro-oculograph (EOG), or electroencephalograph (EEG), to monitor the subject’s alertness. The heart rate or heart rate variability [23] which derived from the ECG signals has been known easily effected by the subject’s psychological and physiological conditions, and therefore the ECG signals is not a good index for monitoring the driver’s alertness. And some laboratories tried to use the electro-oculograph (EOG) signals to define the driver’s alertness. It is reported that the rate of eye blinking [24] was declined along with the decreases of subject’s alertness. However, the time window for analyzing the EOG signals to assess the driver’s drowsiness was around 240 sec, which is too long to use in the drowsiness warning system in the real driving.
Hence, the EEG signals are free from the limitation of long average windows to detect drowsiness. Therefore, EEG remains the most popular modality and the better index used to monitor drowsiness state in real-time.
1.3. Drowsiness related EEG phenomenon of drivers
Previous studies had shown that Along with the subject’s drowsiness level, the neural activities are changed especially in which activities generated from the occipital lobe. Furthermore, the power of occipital alpha (8-12 Hz, [25-29]) and theta band (4-7 Hz, [27-30]) were increased following the decreases of subject’s performances. The similar brain dynamic changes are also observed in a virtual-realty (VR) environment of driving experiments. Lin et al. [31] reported that the power of occipital alpha band was linearly increased from alertness to mild drowsy and then the alpha power was maintain at the same level or slightly decreased from mild drowsiness. In addition, the occipital theta power was also found increased monotonically from alert to deep drowsy. And Lin et al. [32] also demonstrated that EEG is feasible to accurately estimate quantitatively driver’s performance in a realistic simulator by the results above, and constructed 3 editions of EEG monitoring system for drowsiness detection and warning. The first edition [33, 34] was a portable development of wireless brain computer interface using the alpha power increasing in occipital channels to detect drowsiness for warning drivers. Several studies investigated the algorithm for detecting drowsiness by EEG feature. The research team of Lin et al. [26] used independent component analysis (ICA) to remove most of EEG artifacts and suggest an optimal montage to place EEG electrodes for raising average estimation accuracy. Extending previous study, ICA-based fuzzy neural network was used in adaptive EEG-based alertness estimation system for optimizing predict performance [35]. In order to reduce the feature dimension of EEG signals, the nonparametric feature extraction methods were applied to one channel single-trial EEG signal [36, 37]. The latest algorithm reported an unsupervised subject- and session-independent approach for detection departure from alertness [38]. The second edition of EEG monitoring embedded system not only added independent component analysis (ICA) algorithm to monitoring system for raising average accuracy, but also minimized rear-end digital signal
4
processing unit [39]. Duann et al. had shown that it is feasible to correctly estimate the changing level of driving performance using the EEG feature obtained from the forehead non-hair channels [40]. The third edition used the unsupervised algorithm, smaller front-end, cell phone as rear-end and EEG signals from non-hair area without use of gel or skin preparation, and therefore it is more suitable for drivers in practical application [41]. In addition, Lin et al.
investigated the EEG signals changes induced by arousing feedback, and the results shown that significant decrease in power spectra in theta and alpha bands following auditory feedback was found in the bilateral occipital component [42].
The above results suggest that occipital alpha and theta bands would be as good EEG features for indexing the driver’s drowsiness, and the feature extraction algorithms were developed to apply on drowsiness detection and warning embedded system.
1.4. The brain network by Granger causality
The most studies investigating EEG signal in drowsiness state focused on how the EEG power changing along with the subject's driving performance, and it was the only indicator for estimation. In order to obtain more EEG features associated drowsiness for estimation indicator, realizing the brain network from alertness to drowsiness was what we intended to. For comprehending causal relationship between each source of signal from alert to mild drowsiness and to deep drowsiness, one approach to gaining this information is the so-called Granger causality (GC) [43].
Many studies used Granger causality (GC) to analyze EEG or other brain neural signals for realizing the causal relationship between distant brain site [44-46]. The brain networks, causal relationship or neural interactions, means how those signal flows transfer among distinct region in brain under one condition or function. GC could be applied on invasive-recorded local field potential (LPF) [47, 48] functional magnetic resonance image (fMRI) [49, 50] and EEG [51, 52] to construct the brain network. Most research analysis LPF of a small specific brain
area under one experimental function such as the study of Guéguin et al. [46]
used GC to LPF data to investigate the functional connectivity between primary auditory cortex (Heschl's gyrus) and secondary auditory cortex (lateral part of Heschl's gyrus) under amplitude modulated sound processing [47]. However, the location of drowsy cortex is undefined, and therefore LPF is not good signal for observing drowsiness. The fMRI was also used for GC analysis; Duann et al [47]
reported that the great connectivity was generated between inferior frontal cortex and presupplementary motor area during stop signal inhibition by using fMRI data to GC algorithm. Although fMRI data were recorded from whole brain, but the time resolution was not good as EEG data, and therefore it could not reflect the real-time reaction like EEG data in this experiment design. Some studies even applied GC combined with Independent component analysis (ICA). The brain connectivity between the independent components was investigated by applying to the GC analysis of fMRI study of word perception experiment in Londei et al.’s study [49]. The GC analysis in EEG data by Milde et al. [51] has been used for optimizing the adaptive algorithm of GC and realizing the brain network of laser-evoked brain potentials. The EEG signal is more adaptive to the large detectable range and the simulated driving experiment and consequently is the most suitable signal for obtaining the brain networks by GC.
1.5. Aims of this study
The connectivity between independent components at driver’s different drowsy levels was accessed by GC analysis applied in EEG data. The different drowsy levels were defined by the behavior response; the behavior performance could reflect the subject’s consciousness to classify the relative EEG data to different drowsy level. Independent component analysis was used for approaching the signal sources replacing the channel data.
The aims of this study were (1) To determine the concentrations of brain connectivity between different brain regions from subject’s alert status to drowsiness status. (2) To compare the above brain network in motion and
6
motionless simulation, finding the influence of kinesthetic input on EEG signal flows.