In SAE Level 3 automated driving, the driver may not need to continuously monitor the road, but must be able to resume control when the vehicle reaches a situation it cannot handle. A takeover request (TOR) is only useful if the driver can respond in time and intervene safely. Driver takeover prediction offers a way for automated systems to account for that human variability instead of treating every driver as equally ready.
DeepTake is a framework for predicting three related aspects of a driverâs response: whether the driver intends to take over, how long the transition may take, and the quality of the intervention. Its practical significance lies in combining these predictions so an automated vehicle can make decisions with a more informed view of driver readiness.
Why automated vehicle takeover needs a human-centred approach
Drivers may take longer to re-engage with driving after spending time on non-driving-related tasks (NDRTs). Stress and fatigue can also affect their readiness. Even a TOR issued with time to spare does not ensure that a driver will regain control safely. At the same time, frequent alarms may startle drivers and add to stress.
These challenges make a fixed alert strategy difficult: a signal that allows one driver to respond comfortably may not provide enough time for another. Monitoring and predicting driver behaviour could help the vehicle adapt its actions and alerts to the person and situation. The goal is not simply to estimate reaction speed, but to consider whether a driver is likely to respond and how well they may manage the manoeuvre.
DeepTakeâs three dimensions of driver takeover prediction
- Takeover intention: whether the driver is expected to respond to a TOR. In the framework, this is treated as a binary classification task.
- Takeover time: the time between the TOR alarm and the driver beginning to regain control. The framework groups this outcome into three classes, though the modelâs output can be adapted for other class counts.
- Takeover quality: an estimate of how safely the driver intervenes after taking control, also represented as three classes.
Considering all three dimensions provides a broader picture than estimating response time alone. A driver may be likely to respond but still need more time, or may resume control without making an adequate manoeuvre. Separate predictions can help a vehicle distinguish these situations.
How multimodal data supports takeover prediction
DeepTake combines information from several sources: driver biometrics, pre-driving surveys, the type of NDRT being performed, and vehicle data. The rationale is that driver readiness is affected by interacting internal and external factors. Multiple data streams can offer complementary information about a changing state that may be difficult to represent with a single measurement.
Before modelling, the data are pre-processed, segmented and converted into features. Physiological wearable data require particular care because movement can introduce artifacts. The framework therefore calls for pre-processing to remove such artifacts and extract useful information. Features are then labelled according to the relevant takeover outcome.
In the reported study, takeover time was measured from the moment the request alarm sounded until a participant began regaining control by pressing two buttons embedded in the steering wheel. The resulting time was assigned to low, medium or high categories. Takeover quality was assessed in a scenario where the driver needed to steer around an obstacle while avoiding excessive movement from the current lane. The study classified outcomes using lateral deviation: below 3.5 metres as low, from 3.5 to 7 metres as high, and above 7 metres through 10 metres as medium. These thresholds describe the studyâs labelling scheme, rather than a universal standard for takeover quality.
From features to predictive models
DeepTake uses feed-forward deep neural networks (DNNs) to classify takeover intention, time and quality. Each model takes extracted features as input, passes information through successive layers, and produces a classification. The DNNs are trained using mini-batch stochastic gradient descent. The output layer can be configured for multi-class prediction; the framework also demonstrated a five-class setup for takeover time.
Evaluation uses several methods because no single metric captures every aspect of model performance. The approach applies 10-fold cross-validation on training data, compares the proposed model with six other models using receiver operating characteristic analysis and weighted F1 scores, and uses confusion matrices to show how predictions perform across individual classes. These measures offer different views of model behaviour; they do not, by themselves, establish readiness for deployment in a safety-critical vehicle.
Using predictions to adapt vehicle behaviour
In a future integration, takeover predictions could inform when and how an automated system issues a TOR. For example, if the predicted response time is longer than the available time to handle an emerging situation, the vehicle could warn the driver to reduce engagement with an NDRT. If predicted intervention quality is too low, the system could take that into account when determining its next action.
The framework also describes a sequence in which the system infers takeover intention from multiple sensor streams and adjusts vehicle behaviour to suit the driverâs predicted readiness. Where readiness appears insufficient, the system could issue a warning about NDRT engagement rather than assuming the driver can take over appropriately. These are potential uses of the framework, not a claim that DeepTake is a production-ready control system.
The underlying experimental work used simulated driving scenarios on a four-lane rural highway, including five situations in which an automated vehicle might need to request a takeover. The incidents were designed to be unfamiliar to participants, so responses would involve situational awareness and decision-making rather than advance knowledge of when an event would occur. Further validation is important because a framework intended for safety-critical applications must meet relevant safety requirements.
Practical implications for automated vehicle design
DeepTake points toward a more individualised approach to automated vehicle handover. By treating intention, time and quality as distinct prediction tasks, it offers a framework for moving beyond the assumption that all drivers need the same warning and response window. Multimodal sensing and careful data processing are central to that approach, as are clear outcome definitions and evaluation across multiple metrics.
For engineers and system designers, the practical takeaway is that a TOR should not be treated as proof of readiness. Predictions of driver state could help automated systems make more informed decisions about alerts and handover timing, while reducing reliance on a single fixed response assumption. Any real-world application would still require robust validation and safety-focused integration.
Access Official Technical PDF