Stargazer: A Transformer-Based Driver Action Detection System for Intelligent Transportation

Junwei Liang, He Zhu, Enwei Zhang, Jun Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 3160-3167

Distracted driver actions can be dangerous and cause severe accidents. Thus, it is important to detect and eliminate distracted driving behaviors on the road to save lives. To this end, we study driver action detection using videos captured inside the vehicle. We propose Stargazer, an efficient, transformer-based system exploiting rich temporal features about the human behavioral information, with a simple yet effective action temporal localization framework. The core of our system contains an improved version of the multi-scale vision transformer network, which learns a hierarchy of robust representations. We then use a sliding-window classification strategy to facilitate temporal localization of actions-of-interest. The proposed system wins the second place in the Naturalistic Driving Action Recognition of AI City Challenge 2022 (Track 3). The code and models are released.