There are 1.24 million traffic accidents every year with 2.4% caused by drowsy drivers. In that context, several methods for drowsiness detection have been developed. Nevertheless, despite the huge amount of researches, the several devices on markets, and car systems; it is not clear which method is the most appropriate, what sensors are the most useful and least intrusive. This study, shows a systematic literature review of the most recently and relevant methods.
According to the World Health Organization (WHO), 1.24 million traffic
accidents occur every day [
Despite the huge amount of works on the field, and the several numbers of devices for drowsiness detection, it isn’t clear yet which one is the best, which one is the most appropriate according to the conditions of cars and drivers, and what are future about this topic.
This work is a preliminary systematic review of drowsiness detection
techniques. We just take the most recent and relevant works since 2015, we consider
that this work could be the initial step, for persons that want to start their
research on the filed. We take the classification proposed by Ramzan et al. [
We performed our study on search engines such IEEE, ACM, Springer, and Google Scholar. We got 596 research papers; from those, we have selected 265 papers based on the title, those from journals, and the most recently; then 78 papers, were selected after abstract revision; finally, 48 research papers were filtered out as our primary study.
In this section, we included all the non-intrusive methods, they basically relies on sensors embedded on the steering wheel that measures the Steering Wheel Angle (SWA), Steering Wheel Reversal (SWR), Steering Wheel Movements (SWM), Steering Wheel Velocity (SWV), angular velocity, grid force, hand position, absence hands, and Approximate entropy (ApEn). Actually, we could divide the researches in two groups: the cutoff-analysis methods: that measure certain features and use a cutoff in order to detect drowsiness (See Table 3) and machinelearning-based methods that uses machine learning models (See Table 2).
One of the pioneers methods that associated steering wheel data with
drowsiness were Platt in 1963 [
An important aspect in these methods are the frequency they used to log
data. For instance, the SWA is logged with 100Hz [
Moreover, the methods based on steering wheel data are dependant of road
geometry and curvature [
Mθ = 1 nl+w−1
X W k=nl
θ(k) θ∗(k) = θ(k) − Mθ (1) (2) where, W is the length of sliding window, nj , the first point of window and Mθ, the average of steering angle. θ(k) stands for the raw signal and θ∗(k) is the preprocessed signal.
In Fig. 2 (left), is presented the effect of road curvature, the read lines stands for the raw data and the blue lines represent the preprocessed data (after curvature effect removal). As we can see, the read curve have spikes that represent the rad curvature. 3.2
In this section, we include intrusive methods, they basically measures the Heart
Rate (HR), Heart Rate Variability (HRV), Pulse Rate (PR), Breathing Rate
(BR), Respiratory Rate (RR), body temperature, electrical brain activity and
electrical eye activity. The device used are electroencephalogram (EEG),
Electrooculogram (EOG), Electromyogram (EMG) and electrocardiogram (ECG).
These methods are more reliable and accurate [
The EEG is the most common device, a EEG have 7 bands representing the
state of brain, the bands most used are α, β and θ. Moreover, these methods
could be divide in FFT-based spectral analysis, wavelet-based spectral analysis
and Higher Order Statistics-based analysis [
Other methods use the Hearth Rate data [
Behavioral methods are based on image processing of drivers capture with a
camera. Features as eyes, mouth, head pose, percentage of eye closure
(PERCLOS), face, eye blink, eye closure are used, the majority uses machine learning
classifiers. These methods are non-intrusive and depend a lot from
illumination and the image quality but with the introduction of deep learning and best
camera devices, these methods become promising. Moreover, databases are used
for both, image and video processing. Some public databases are DROZY [
In this approach, almost every research have been built its own dataset [
In addition, there is not a comparison between all methods but the most
prominent works, could be Haruna et al. [
After analysis, it is appreciated that SVM is the most widely used classifier, as
it provides mostly greater precision and speed, but it does not work well for large
data sets, in Table 6 we present the most relevant methods. On the other hand,
both CNN and HMM are slow in training and expensive. The methods that have
the highest percentage of certainty are those that work with deep learning, and
most of them use CNN,since it uses special convolution and pooling operations
and performs parameter sharing. This enables CNN models to run on any device,
making them universally attractive. The majority of works carried out tests for
the detection of drowsiness with databases in simulated environments, and the
lesser number of works that carried out tests in an uncontrolled environment,
which required gathering more than one method or combining characteristics to
achieve the detection with less margin of error, the work that proved to be most
relevant with deep learning is Babitha et al. [
The state of art methods for drowsiness detection was reviewed in this study. Three main approaches were analyzed: behavioral approach, based on the analysis of image and video of drivers; vehicular approach, based on devices inside the vehicle, the majority are sensors embedded on the steering wheel; and the physiological approach, these intrusive methods are devices that a driver has to use on the head, hands, and fingers.
Some authors claim that the physiological methods have the highest accuracy but there is no clear evidence of a comparison between them Moreover, deep learning is used intensively, but SVMs and features vectors are still used.
The main disadvantage in this field is related to the dataset. The researches are not releasing them. Moreover, and more alarming, the few public datasets, are not in real conditions, they are recorded in offices or are simulated, as opposed to having your own training data from a larger sample of participants, while including new distinct signals of drowsiness (sudden head movement, hand movement, or even tracking eye movements, and others).