Fused Architecture with Enhanced Bag of Visual Words for Efficient Drowsiness Detection

Drowsy driving is more hazardous than reckless driving. This study concentrates on capturing the behavioral features of drowsiness from facial images of a driver. The methodology considers scale invariant feature transform matched with the fast library for approximate nearest neighbors for low-level drowsy features extraction. These features are fused with the high-level features extracted from the convolutional layers of a convolutional neural network (CNN). The convolution operation incorporates a model parallelization technique to increase the efficiency of the training and improve the feature identification. Further classification is performed by considering the occurrences of visual words using the softmax layers of the CNN. In contrast to existing state-of-the-art models which require a few seconds to detect drowsiness, this model detects drowsiness in milliseconds. With the model parallelization approach, this model exhibits a high accuracy rate of 83.8% relative to normal CNNs.