Day Driving and Night Driving Behavior Detection Using Deep Learning Models

The detection of driving behavior is critical for road safety and driver assistance systems. Comparing the performance of driving behavior detection during day and night provides insight into the strengths and weaknesses of these systems under different lighting conditions. This comparison can encourage further work to improve the accuracy and reliability of the systems and ensure they always function optimally. This study aims to compare the performance of driving behavior detection using a digital video recorder (DVR) camera during day and night driving. The study uses a dataset of video footage recorded by a DVR camera on a driving simulator during both day and night driving. Several Convolutional Neural Network (CNN) models have been developed to detect instances of distracted driving behaviors. The study, conducted at VITAL Laboratories at UiTM Shah Alam, used MobileNetV2, GoogleNet, InceptionV3, and Resnet50 to model the data from 44 participants showing 7 different driving behaviors: normal driving, yawning, nodding off, texting, calling, talking to passengers, and using car features. Based on the experiments, the best CNN model for both sessions is InceptionV3 with the overall accuracy for driving behavior detection during day driving is 92.2% and during night-time is 76.9%. The results show that driving behavior detection performs better during the day driving session than at night driving session by a difference of 15.3%.