Real-Time Detection of Distracted Drivers Using a Deep Neural Network and Multi-threading

Abstract

Convolutional neural network (CNN) is a very popular deep learning architecture used for the problem of object detection and classification of images and videos. CNN is appropriate for image detection as it prunes the computational overheads as well as provides better performances. Visual Geometry Group (VGG) model is an instance of CNN architecture to solve the problem of object detection from images. VGG-16 model is one of the most potent classifiers that had won the second place in the ImageNet ILSVRC-2014 challenge. VGG also helps in reducing latency of detection of distracted drivers in real time (RT). The technique of transfer learning was applied using VGG-16, pre-trained on the ImageNet dataset intended to extract bottleneck feature that is further used to train a classifier. This technique thrives on the combined power of VGG convolutions and multi-layer perceptron. A key highlight of this paper is the coupling of transfer learning using VGG-16 with multi-threading to achieve a real-time prediction of distraction, from a video input. The proposed model performed notably well on real-time video, achieving an accuracy of 96%, and a video output of 26 frames per second which is comparable to the state-of-the-art algorithms, for real-time classification of objects.