Identification of dangerous driving state based on lightweight deep learning model

Abstract

Dangerous driving state is a significant cause of traffic accidents and traffic safety problems., and the relevant detection means are lacking at present. In order to monitor the status of drivers in real time, timely judge the dangerous driving time and give reminders, this paper proposes a hazardous driving image classification system based on a modified ShuffleNet lightweight model to deploy it flexibly inside a vehicle or in a particular detection device, which is trained, tested and applied on a professionally collected image dataset. The experimental results demonstrate that the proposed system can effectively reduce the model complexity, increase the speed of model operations and usage without significantly affecting the classification effect through comparing with traditional deep learning models. Various visualization schemes are used to analyze and compare the image feature information extracted by the various models to investigate the feature extraction effectiveness of lightweight deep learning image classification models.