Video-based Driver Distraction Detection using Transfer Learning: Bangladesh Perspective

Bangladesh faces a severe challenge with a high frequency of road accidents, often caused by distracted drivers. Driver distraction is a significant contributor to road accidents and accounts for a large number of road crashes and injuries. Many researchers have experimented with various deep learning approaches in simulated environments to automatically address this situation worldwide. Unlike other studies, we use real-world images from Bangladesh, creating a dataset that fits the local context. We evaluated five deep learning models using transfer learning to detect driver distraction on this new dataset. Performance comparisons are provided by using five recent deep learning models: MobileNetV2, DenseNet201, ResNet50, InceptionResNetV2, and InceptionV3. The experiment shows that without using data augmentation, InceptionV3 achieved an accuracy of 75%. On the other hand, with data augmentation, both MobileNetV2 and DenseNet201 showed an accuracy of 72%.