AI Insight
This systematic review analyzes 85 studies on driver state monitoring in SAE Level 3 automated vehicles, where the system controls driving but requires drivers to resume control when needed. The review examines how driver states are defined and measured, what sensing technologies are used, and what computational methods including multimodal fusion detect driver readiness for takeover transitions. The authors identify key gaps including lack of standardized datasets, inconsistent operational definitions, unrepresentative sampling, and insufficient real-world longitudinal validation.
Why it matters
Safe implementation of Level 3 automated driving systems depends critically on accurate real-time assessment of whether drivers can safely resume control. This review provides a comprehensive framework for developing and evaluating driver monitoring systems, highlighting technical and methodological priorities needed before widespread deployment of conditional automation.
Understand the Science
by Haoran Wu, Timothy Tettey Nartey, Guangjun Wan, Yugang Wang, Linli Xu, Xun Zhou
SAE Level 3 automated driving systems assume full control of the dynamic driving task within their operational design domain. Ensuring a safe and timely transition from automated to manual driving therefore necessitates continuous monitoring and assessment of driver state. This review presents a synthesis of 85 studies published mainly between 2021 and 2025. A PRISMA 2020-guided search was conducted in ScienceDirect, Web of Science, SpringerLink and IEEE Xplore, with the final search completed on 15th September 2025. Only English-language publications that examined driver-state detection, recognition or assessment in driving scenarios relevant to conditional automation were considered. Studies published before 2021 were excluded, unless they provided foundational contributions. This review was not registered. This review aims to (1) identify driver-state constructs that have been investigated in conditional automated driving, examining how they are defined and operationalized; (2) characterize sensing modalities used for driver-state monitoring; (3) evaluate computational approaches including multimodal fusion used in previous studies; (4) examine performance metrics, validation strategies and real-time feasibility; (5) identify methodological limitations, reporting inconsistencies and priorities for future development and deployment. This study suggests the standardisation of Level 3 datasets. It also calls for harmonised operational definitions and more representative sampling. It finally advises the performance of more longitudinal on-road validation. This work was supported by the National Natural Science Foundation of China.
Source: Driver state detection and recognition in conditional automated driving: A review