Driver Distracted Behavior Detection Using a Light Weight Model based on the W-MSA

Abstract

The percentage of traffic accidents caused by driver factors is about 90% in the world. Despite the great development of autonomous driving, it is still not completely self-driving. So, it is still not possible to avoid traffic accidents caused by drivers. Computer vision technology has made great progress with deep learning development. That makes it possible to detect the driver’s behaviour using a camera. To reduce the detection price, this paper presents a light weight model to detect the driver’s behavior based on the W-MSA. This model consists of 2 encoder modules and a classification module. And it used the Global Avgpool and W-MSA to reduce the model parameter and FLOPs. To avoid the low accuracy of the detection, this paper also used label smoothing regularization and CBAM technologies to improve the accuracy. This paper also used a visualization method to show the interpretability of the proposed model. The results show that the accuracy of the proposed model is 98% on the Kaggle driving test dataset. Compared to other state-of-the-art models, our method has a high accuracy with fewer model parameters.