Tuning of Hyperparameters and CNN Architecture to Detect Phone Usage During Driving

The primary goal of this paper is to classify images using Convolutional Neural Networks (CNN) to determine whether a driver is using their phone while driving or is paying attention to the road. There are a lot of road accidents happening on a day-to-day basis for which one of the main causes was found to be the driver behaviour. Talking on or texting on a smartphone while driving, along with other distractions, can be highly dangerous. Recognizing these poor driving habits can prevent many unfortunate events as well as save many lives. In this study, we propose a CNN model to categorize photos of drivers on a binary basis to ascertain whether the driver is using a phone while driving or is driving safely. Dataset was created using images from Google to conduct the analysis. For model training, equal proportions of the two image categories—driving safely and talking on the phone—were taken. Customized CNN models were created using different layers, and their accuracies were tested. Manual hyper parameter tuning was done on various parameters as well as the CNN architecture. The best model among all the models was the one which had 2 layers of convolution before 1 layer of pooling, along with dropout layers and was trained with 20 epochs. This model can be used in automobiles or in the traffic system to maximize the safety on the road.