Table 5.

Summary of the technologies discussed in this paper. NA means that the accuracy could not be ascertained within the reference.

Methods	Measurement Information	Previous Studies	Method	Accuracy
Contact method	Biometric information	Satti et al. [38]	Electromyogram measurement from electrodes attached to the steering wheel Electrocardiographic measurements from a wearable sensor on the wrist.	NA
		Kundinger et al. [42]		≧92%
		Kundinger et al. [43]		≧90%
Non-contact method	Vehicle behavior	Subaru [44] Hino [45] Mazda [46] Honda [47] Volvo [48] Jaguar [49]	Detects changes in vehicle behavior and warns from HMI.	NA
		Arefnezhad et al. [10]	Apply ANFIS with steering angle as input.	98.12%
		Jeon et al. [50]	Estimation by ensemble network model using steering and pedal pressure as input.	94.2%
	Graphic information (of driver)	Toyota [51] Subaru [53] Nissan [54] Hino [45] Thanko [55] Yupiteru [56]	Warnings for closed eyes and side glances.	NA
		Toyota [52]	Stops the car when the driver is not in a good posture or does not respond to warnings.
		Cardone et al. [61]	Applied PERCLOS to visible images obtained by a thermal imaging camera and classified “wakefulness”, “fatigue”, and “dozing” by deep learning. Support vector machine, K-nearest neighbor method, and decision tree were used to classify sleepiness based on the temperature patterns of the forehead and cheeks.	Approximately 65%
		Tashakori et al. [63]		84%
Non-contact method	Graphic information (of driver)	Celecia et al. [62]	Fuzzy inference system to estimate sleepiness from eye and mouth information.	95.5%
		Chakkravarthy [64]	EAR	75% when blinking, 35% when wearing glasses, and 25% when hair is hanging over the face
		Manu [67]	Correlation coefficient template matching.	94.58%
		Li et al. [69]	Detecting fatigue from driver’s eye closure time, few blinks, and few yawns.	95.10%
		Képešiová et al. [70]	Learning grayscale face images with CNN.	98.02%
		Dua et al. [71]	Detects drowsiness by considering four different types of features (hand gestures, facial expressions, behavioral features, and head movements) using four deep learning models: AlexNet, VGG-FaceNet, FlowImageNet, and ResNet.	85%
		Yang et al. [72]	Nodding detection using LSTM autoencoder on RFID tag data.	≧90%
		Jabber et al. [74]	Facial landmarks from images were detected and estimated by a system based on multilayers perception classifiers.	81%
		Ma et al. [75]	Classified the driver’s drowsiness by PSO-H-ELM based on the power spectrum density of EEG data.	83.12%
Multiple methods		de Naurois et al. [6]	Modeled using the information on eyelid closure, eye and head movements, and driving time. Logistic regression with Eye Closure, head movement, KSS, HFC, etc., as explanatory variables.	MSE of drowsiness level: 0.22
		Baccour et al. [27]	Pulse, respiration, and center of gravity information were obtained, and ESN was used for estimation.	72.7%
		Ariizumi et al. [76]		83.3%