Previous Research

Drowsiness leads to decline in drivers’ abilities of perception, recognition, and vehicle control and hence monitoring of drowsiness in derivers is very important to avoid road accidents [9]. Some researches used non-physiological method, as eye closure with CCD image tracking [10]-[16]. And others used physiological parameters to increase the accuracy of drowsy detection, like pulse wave analysis with neural network [20], the electrooculography (EOG) and the electromyography (EMG) measurement [17], [18], and theelectroencephalogram (EEG) [19]-[21].

In 2003, Hamada et al. proposed a driver status monitor system by using CCD camera, as shown in Fig. 1-1 [13]. The CCD camera was installed in the car and focused on the user’s eyes. The driver status monitor detected drowsiness from the change in the duration of eye closure during blinking and inattention from the change in the gaze direction. Using CCD camera to contribute the urgency system was a very difficult work here. There were some critical points inside, and needed to overcome.

For instance, user couldn’t move for free, the images detecting performance were easily be interfered by light, and the largest problem was that the system is too big, complex, and expensive to implement. The algorithm of eye tracking also needed to

use edge detecting to train data, and hence to build up a neural network to classify the drowsy status.

Fig. 1-1: The role of driver status monitor [13]

An alternate is to detect the moment from alertness to drowsiness by using physiological parameters. In 2005, Thum et al. used EOG as an alternative to video-based systems in detecting eye activities caused by drowsiness [18]. Rapid eye movements (REM), which occurred when one is awake, and slow eye movements (SEM), which occurred when one is drowsy, can be detected through EOG. The results showed that the detection rate for eye activities caused by drowsiness was more than 80 %. However, REM and SEM are difficult to measure when users are driving because users can not close his/her eyes when they are driving a vehicle on the road, and then SEM is hard to measure. In addition, REM and SEM are tending to the level of sleep stage not the indicator of drowsiness detection, so they can not be used as the parameters of on-line process.

In 2003, Caffier et al. proposed that the spontaneous eye blink is considered to be a suitable ocular indicator for fatigue diagnostics [24]. To evaluate eye blink parameters as a drowsiness indicator, they developed a contact free method for the measurement of eye blinks by using an infrared sensor clipped to an eyeglass frame recorded eyelid movements continuously. The parameters blink duration and

reopening time in particular change reliably with increasing drowsiness. The results demonstrate that the measurement of eye blink parameters provided reliable information about drowsiness. In 2008, Jammes et al. in order to automatically score the drowsiness level, they developed a software for identifying blinks in EOGs as their first step [23]. They recorded vertical EOG signals by surface electrodes placed above and below the eyes. The analysis of EOG velocity based on expert rules was the originality of their blink detection algorithm and more than 97.7% of blinks were detected by their algorithm. The drowsiness scale they selected was Karolinska Drowsiness Score (KDS) which would score when signs of drowsiness, i.e. long duration or small amplitude blinks were detected. Comparing the results of KDS and the results of their automatic scoring, and then they found out the correlation of these results was high. It demonstrated that blink duration and amplitude are important parameters for drowsiness detection.

Brain Computer Interface (BCI) is an interface between human and computers or machines. It is based on the translation of the specific brain activity generated by a specific thought of a human to control machines, to communicate with the outside world directly, to convey the message, and independent operations, as well as self-care purposes. BCI can be divided into three distinct modes: invasive, partially-invasive, and non-invasive BCI. Non-invasive BCI is the main stream of BCI research which has advantages of both easy application and absence of procedural risks, such as infection or cortical micro-lesions. There are several approaches to non-invasively acquire brain activities, such as magentoencephalography (MEG), positron emission tomography (PET), functional magnetic resonance imaging (fMRI), electroencephalography (EEG) and et al. EEG is the mainstream of non-invasive BCI, because of its much fine temporal resolution,

ease of use, portability and low set-up cost. In particular, higher temporal resolution becomes the great temptation to use EEG techniques as a direct communication channel from the brain to the real world [27]-[42].

In EEG system, it was different from other physiological parameters, and moreover it owned intuitive and specific characteristics, such as alpha, theta or beta band power followed subject’s own mental state. In addition, the EEG system usually needed to collect enough EEG data to analyze. The supervised methods which previously study often had been used to train a learning data, and usually implement in off-line EEG analysis. Previous studies which used supervised methods developed several kinds of brain computer interface for drowsiness detection [19], [20]. When the subject changed the state from alertness to drowsiness, the alpha rhythm will increase and beta rhythm will decrease [21]. In 2005, a drowsy estimation system was developed by combining independent component analysis (ICA), power-spectrum analysis, correlation evaluations, and linear regression model to estimate a driver’s cognitive state when he/she drove a car in a virtual reality (VR)-based dynamic simulator [19]. Its flowchart of EEG processing was shown in Fig. 1-2. In the above studies, an EEG machine, Scan NuAmps Express system (Compumedics Ltd., VIC, Australia), was used to measure EEG, as shown in Fig. 1-3. It is not small, light, and wearable. Moreover, the above algorithms for drowsiness detection requires mass computation complexity, thus, they are not easy to be implemented in a portable DSP device.

Fig. 1-2: Flowchart of EEG processing in drowsy estimation system [19]

Fig. 1-3: Scan NuAmps Express system (Compumedics Ltd., VIC, Australia) In the supervised mode, supervised learning methods such as artificial neural network (ANN) could be used to classify different states of vigilance. But stimulus may introduce some noise. So in [43], the author proposed a semi-supervised learning algorithm which can quickly label huge amount of data. Here another author proposed another kind of semi-supervised learning method based on probabilistic principle component analysis (PPCA) to distinguish wake, drowsy and sleep in driving simulation experiment. After training with data of around 20 min (6–8 min for each state), they could directly use our method as a real time classifier to estimate driver’s

vigilance state [44]. Although this method could greatly reduce the training time, but it still must used in off-line analysis. In our target, we wanted to find a non-training and unsupervised method, and easily implement to an on-line detecting system.