Individual variability in physiological features of drowsiness poses challenges for reliable detection. This study addresses these issues by proposing a multimodal sensor system that utilizes a variety of signals to improve reliability and adaptability to individual subject characteristics. The human drowsiness detection is performed by the fusion of model outputs for the three signal sources. Electroencephalogram signal from frontal (Fp1, Fp2) and occipital (O1, O2) electrodes and facial video of the subject are considered as input data. The shape of blink artifacts is analyzed in the signal from the frontal electrodes, and the presence of increased alpha activity is analyzed in the signal from the occipital electrodes. The eye segment states are classified from video. The dataset for training and testing neural network models is composed of the records of experiments of 5 subjects. Each subject has individual characteristics of physiological parameters, such as weak or strong expression of alpha rhythm and blink artifacts. Convolutional neural networks, autoencoders and recurrent neural networks have been applied to process these signal types. An accuracy of 90.7 ± 3% was obtained as a testing result of the two-level classifier.