Abstract
In open traffic environments, humans still have to remain in the control loop of vehicles due to the limitations of existing technologies and high costs. Realizing effective cooperation between human drivers and automatic systems requires precise timing for transitions. This paper proposes a driving capability-based transition strategy that considers both driver correction ability and driving risk. The transition point is determined by an "unreliable domain" (UD), with boundaries modeled using a log-normal distribution and updated online via an adaptive algorithm. Validation on a driving simulator shows that this strategy accurately identifies transition times, reserves sufficient time for system correction, and adapts to individual driving characteristics.
Introduction
Cooperative driving aims to combine human intuition with machine precision. A critical challenge is determining the "switch point"—when the vehicle should take control from the human driver. Existing methods often rely on continuous shared control, which can lead to skill atrophy, or state-based switching using intrusive cameras to detect fatigue. These can lead to "nuisance" interventions if a driver is momentarily distracted but still in control. This research addresses the need for a non-intrusive, personalized system that intervenes only when a driver is objectively incapable of managing current risks.
Methodology: The Unreliable Domain
The proposed transition logic rests on two pillars: Driving Risk and Correction Ability. Driving risk is measured using standard metrics like Inverse Time-to-Collision (TTCi) for longitudinal safety and Time-to-Line Crossing (TLC) for lateral lane-keeping.
Correction ability is quantified by comparing a driver's actual steering and acceleration inputs to a "desired" command predicted by an Artificial Neural Network (ANN-NARX) model. This model acts as a "shadow driver," representing how a competent driver would behave in the same scenario.
These factors define a 2D space containing an Unreliable Domain (UD). If the combination of high risk and low correction ability falls within this domain, the transition logic triggers the automatic system. To account for individual differences, a recursive algorithm updates the UD boundaries (modeled as a log-normal distribution) in real-time, allowing the system to learn a specific driver's style within approximately 15 minutes.
Experimental Setup and Data Collection
The strategy was validated using a high-fidelity "driver-in-the-loop" platform. The setup integrated Prescan and dSPACE (MicroLab/ASM) with physical actuators to simulate realistic driving conditions.
Data was collected from ten participants who performed visual-manual distraction tasks, such as solving math problems on a phone, while driving. This provided a dataset of "incapable" driving states to test the system's detection accuracy. The simulator's realism was further validated against 30 km of naturalistic driving data from Chongqing, China.
Results and Key Insights
The UD-based strategy demonstrated high accuracy in identifying risks, catching 91.3% of lane departure risks and 88.4% of rear-end collision risks. Notably, the inclusion of "correction ability" as a metric resulted in zero false positives—no unnecessary interventions occurred during testing.
Key insights from the study include:
- Capability Over State: Distraction does not always equate to incapability. By focusing on deviations from predicted ideal control commands rather than biological markers (like eye tracking), the system avoids unnecessary interventions during safe moments.
- Personalization: The adaptive log-normal distribution ensures that the system's threshold for intervention is personalized, recognizing that what is risky for a novice might be standard for an expert.
Conclusion and Future Work
The proposed strategy offers a robust, non-intrusive method for managing handovers in semi-autonomous vehicles. While successful in a simulator environment, future work is needed to test the model against the "noise" of real-world traffic. Additionally, expanding the scope beyond visual-manual distraction to include fatigue or emotional distress, and exploring graded "haptic shared control" instead of binary transitions, are promising avenues for further research.
Summary updated