Non-intrusive drowsiness detection is an important design goal for driver monitoring systems: the system must recognize declining alertness without requiring the driver to wear sensors or make contact with measurement equipment. Camera-based behavioral monitoring offers one route. Research using naturalistic driving data found that video-derived features could classify drowsiness with performance similar to an electrooculogram (EOG) signal, while combining video and electrocardiogram (ECG) features produced higher accuracy in the tested binary classification task.
For engineers and technical decision-makers, the results point to a useful balance between measurement quality and real-world practicality. Physiological signals can be informative, but their acquisition may be intrusive. Cameras are easier to integrate into a vehicle environment, yet their performance depends on reliably extracting useful features under changing conditions.
Why non-intrusive drowsiness detection matters
Drowsiness reduces vigilance and cognitive performance, creating safety risks during driving. It can result from sleep-related factors, such as sleep restriction and circadian rhythms, or from the driving task itself. Monotonous conditions may contribute to passive fatigue, while demanding conditions can create task-related strain. The causes differ, but reduced alertness can impair driving performance.
Detection approaches generally use subjective ratings, vehicle behavior, driver behavior, physiological signals, or combinations of these. Vehicle-based methods can monitor indicators such as steering movement or lane position, but they may be affected by road geometry, conditions, vehicle characteristics, and other forms of distraction. Physiological measurements can provide useful signals, but commonly require sensors attached to the person. Behavioral systems instead analyze observable cues, including eye and head movement, using a camera.
What a camera can measure
Video-based driver monitoring turns facial and eye movement into measurable features for a classification system. The study evaluated features related to head orientation, eyelid opening, pupil diameter, and gaze direction. These measures can help represent changes in how a driver looks, moves, and maintains eye opening as alertness changes.
Eye-related indicators include blink characteristics and the duration or proportion of time the eyes remain closed. PERCLOS, for example, describes the proportion of time that the eyes are closed during a defined period. Head orientation and gaze add information about where the driver is looking and how their head is positioned. No single feature needs to represent the entire state: a machine learning model can use several features together to distinguish alert and drowsy periods.
A camera-based approach avoids the need to attach an EOG sensor to the driver, which makes it relevant to practical in-vehicle applications. However, the camera still needs to capture the face well enough for those features to be extracted. Lighting conditions and camera positioning remain important technical considerations.
Comparing video features with physiological signals
The study applied supervised machine learning classification to measures from SleepEye, a validated dataset collected during a naturalistic field study on real roads. The available data included previously extracted video measures, so the analysis compared these features with physiological signals rather than processing the original video for every experiment.
Using 17 video features, the best reported results were approximately 84% accuracy for a two-class task—awake versus drowsy—and about 73% for a three-class task—awake, medium, and drowsy. The video-based results were similar to those obtained using EOG alone. This comparison suggests that camera-derived behavioral measures can be a viable alternative when intrusive physiological acquisition is undesirable, while also showing that the classification task becomes more challenging when the system must distinguish three alertness levels rather than two.
These results describe performance on the studied dataset and selected features. They should not be treated as a guarantee of equivalent performance in every vehicle or operating condition.
Hybrid sensing and machine learning
Behavioral and physiological signals capture different aspects of a driver’s state. Combining them may help address the limitations of relying on a single measurement type. In the study, video features were combined with ECG features in a hybrid classification approach. The best reported accuracy was close to 89% for the binary problem. As with the video-only experiments, performance on the three-class task was roughly 10 percentage points lower than on the binary task.
Several additional strategies were examined, including classifier combinations, approaches to imbalanced data, and methods for sequential data. These tests did not produce considerable improvements. The findings indicate that the tested feature set and dataset may limit further gains from changing the classification strategy alone.
For system designers, hybrid sensing is therefore a promising direction, not an automatic solution. A design must weigh the potential value of combining information against the practical requirements of acquiring each signal. If the goal is a non-intrusive installation, video measures may be the central input; where physiological data can be collected, a hybrid approach may offer additional information.
Deployment considerations for camera-based systems
A separate experimental framework in the work was developed to extract drowsiness-related measures from a driver’s face using a standard camera. The purpose was to support richer feature extraction than was possible with the pre-extracted measures available in the dataset. The approach demonstrates a path toward camera-based monitoring, but further work is needed to make the algorithm more robust to different lighting conditions and camera positions.
These constraints matter when moving from model evaluation to a deployed driver monitoring system. Engineers should consider whether the camera consistently captures the face and whether the selected features remain available as the viewing conditions change. They should also distinguish between binary detection and graded alertness classification: the reported results show the latter is more difficult.
Practical implications
Non-intrusive drowsiness detection offers a practical way to monitor driver behavior without relying solely on contact-based physiological sensors. In the evaluated naturalistic dataset, video features performed similarly to EOG alone, and combining video with ECG improved the best binary accuracy to nearly 89%. At the same time, lower three-class accuracy and unresolved sensitivity to lighting and camera placement highlight the limits of the reported work.
For technical teams, the key takeaway is to treat sensing and classification as connected design decisions. Camera-derived features can support real-world-oriented monitoring, while hybrid methods may add value when additional signals are feasible. Robust feature extraction across operating conditions remains essential before these approaches can be relied on in broader applications.