Distracted driving causes many accidents every year, most of which can be avoided with automatic recognition. As a result, vision-based driver action recognition is receiving increasing research attention. In a limited in-vehicle space, actions can be very ambiguous from an individual view. Therefore exploring efficient multi-view action recognition architecture is meaningful. This study aims to detect the distraction of drivers while identifying the cause. A novel driver action recognition architecture named multi-view vision transformer (MVVT) is proposed, which combines classical convolutional neural networks (CNNs) with vision transformer. Self-attention mechanism is utilized to dynamically aggregate temporal information and fuse features of different views jointly. Experiments demonstrate that MVVT can effectively recognize drivers’ behaviors with multi-view input. A promising result of 84.9% accuracy is achieved on a large public driver action dataset.
Multi-view Vision Transformer for Driver Action Recognition
Lect. Notes Electrical Eng.
International Conference on Intelligent Transportation Engineering ; 2021 ; Beijing, China October 29, 2021 - October 31, 2021
2021 6th International Conference on Intelligent Transportation Engineering (ICITE 2021) ; Chapter : 85 ; 970-981
2022-06-01
12 pages
Article/Chapter (Book)
Electronic Resource
English
Multi-view Vision Transformer for Driver Action Recognition
TIBKAT | 2022
|Multi-view Vision Transformer for Driver Action Recognition
British Library Conference Proceedings | 2022
|Driver-Skeleton: A Dataset for Driver Action Recognition
IEEE | 2021
|