Driver Fatigue Monitoring: Review and Insights

Driver Fatigue Monitoring: Review and Insights

Setu Shankhdhar

Department of AIML Moradabad Institute Of Technology

Moradabad, India setushankhdhar95@gmail.com

Bhumika Pandey

Department of AIML

Dr. Hina Hashmi

Department of AIML Moradabad Institute Of Technology

Moradabad, India

Geeta Gerola

Department of AIML Moradabad Institute Of Technology

Moradabad, India geetagerola564@gmail.com

Pushpanjali Sharma

Department of AIML

Moradabad Institute Of Technology Moradabad, India pandeybhumika0406@gmail.com

Moradabad Institute Of Technology Moradabad, India pushpanjalisharma796@gmail.com

Abstract-Driver drowsiness has emerged as a critical cause of road accidents, resulting in thousands of preventable injuries and fatalities worldwide. Drowsiness refers to a reduced state of alertness where a driver struggles to stay awake, often leading to unintentional micro-sleep episodes. This review paper presents a comprehensive survey of existing driver drowsiness detection systems, focusing on three major categories: physiological-signal- based methods, facial-feature-based approaches, and vehicle- driving-behavior analysis. Recent advancements in each cate- gory are examined in detail, highlighting the underlying tech- niques, algorithms, and implementation strategies proposed in contemporary research. A comparative evaluation of recently published studies is also provided, considering factors such as accuracy, robustness, hardware dependency, cost, and level of intrusiveness. The advantages and drawbacks of each approach are systematically discussed. Findings from the review indicate that no single method is universally optimal; however, hybrid approaches that integrate multiple modalities demonstrate higher reliability and better real-time performance. Such multimodal systems hold strong potential for developing efficient, practical, and highly accurate drowsiness detection solutions for modern intelligent transportation systems.

Index Terms-Driver Drowsiness Detection, Physiological Sig- nals, Facial Feature Analysis, Vehicle Behavior Monitoring, Intelligent Transportation Systems.

1. INTRODUCTION

   Road safety studies consistently highlight drowsy driving as a major contributor to traffic accidents worldwide [1], [2]. According to the National Highway Traffic Safety Admin- istration (NHTSA), thousands of crashes and fatalities each year are linked to drivers who become fatigued while behind the wheel [3]. Moreover, researchers indicate that the actual number of drowsiness-related accidents is significantly under- reported, suggesting that the real prevalence may be several times higher than official records [4].

   Drowsiness is defined as a state in which an individual struggles to remain awake or alert, often drifting toward sleep even during active tasks such as driving [5]. This condition is strongly influenced by the human circadian rhythm, which

   Identify applicable funding agency here. If none, delete this.

   naturally lowers alertness during specific hours-especially between midnight and early morning [6]. During these periods, reaction times slow, focus decreases, and the likelihood of micro-sleep episodes increases, making driving particularly hazardous [7].

   Driver drowsiness can be identified through behavioral and physiological indicators. Behavioral symptoms may include delayed reactions, poor coordination, lane deviation, and dif- ficulty maintaining a consistent speed [8]. Physiological signs often appear on the driver's face, such as frequent blinking, prolonged eye closure, yawning, head nodding, neck stiffness, or short episodes of micro-sleep [9]. Although these signs are clear warning signals, drivers commonly underestimate their fatigue levels and continue driving without taking necessary breaks [10].

   To understand real-world patterns of sleepy driving, several population-based studies have been conducted. For instance, a large-scale survey in Belgium assessed self-reported drowsi- ness levels using the Karolinska Sleepiness Scale (KSS) [11]. The study revealed that a notable percentage of trips were made by drivers experiencing moderate to severe drowsiness, emphasizing the need for effective monitoring and early de- tection systems.

   Due to the severe risks associated with fatigued driving, researchers have developed numerous techniques to automat- ically detect drowsiness. These methods generally fall into three major categories [12]:

   1. Physiological-signal-based approaches – analyzing signals such as EEG, ECG, EOG, or heart rate vari- ability.

   2. Facial-feature-based approaches – monitoring visual cues like eye closure, blinking rate, yawning, and head movements.

   3. Driving-pattern-based approaches – evaluating devi- ations in steering behavior, braking patterns, lane posi- tion, and vehicle speed.

   Each technique offers unique advantages but also comes with limitations related to cost, hardware requirements, intru-

   siveness, accuracy, and real-time performance. This review pa- per aims to provide an in-depth overview of these techniques, compare their effectiveness, and highlight their strengths and weaknesses based on recent research.

   The remaining sections of this paper are structured as follows:

   • Section II presents a detailed review of existing drowsi- ness detection methods across the three major categories.

   • Section III provides a comparative analysis and discusses observations from recent studies.

   • Section IV concludes the paper and outlines potential future research directions, particularly the development of hybrid systems that integrate multiple detection tech- niques for enhanced accuracy and reliability.

2. LITERATURE REVIEW

   Research on driver drowsiness detection has evolved sig- nificantly over the past two decades, shifting from manual observation to sophisticated automated systems. Early studies primarily relied on physiological sensors to capture real-time biological responses that correlate with fatigue. However, as technology advanced, computer vision and machine learning became dominant approaches due to their non-intrusive nature and ability to work in real time. This section reviews major methodologies proposed in literature across three primary domains: physiological signals, facial features, and driving behavior.

   In recent years, physiological-signal-based detection meth- ods such as EEG, ECG, EOG, and heart-rate variability (HRV) have shown strong reliability in identifying fatigue at early stages by directly monitoring the brain or cardiac activity. These techniques measure variations in neurological or cardiovascular responses that occur during drowsiness.

   1. Physiological-Signal-Based Approaches

      One of the earliest and most reliable approaches involves physiological-signal analysis, particularly the use of EEG (electroencephalogram) to measure brainwave activity. Re- searchers such as Makeig et al. [1] demonstrated that tran- sitions from alert to drowsy states could be detected by monitoring variations in alpha and theta waves. Their system provided high accuracy, but practical deployment was limited because EEG headsets are intrusive and uncomfortable for everyday drivers. Later studies attempted to use dry electrodes, but reliability still decreased under real driving conditions.

      Another physiological approach focuses on heart rate vari- ability (HRV) and ECG signals. Vicente et al. [2] developed an ECG-based fatigue detection system using time-domain HRV features. While this method showed promising results in early stages of drowsiness, it suffered from noise interference caused by driver movement and required wearable devices. These limitations made physiological methods accurate but not user- friendly for long-term driving.

   2. Facial-Feature-Based Approaches

      With the rise of computer vision, researchers shifted toward facial-feature-based techniques, which rely on camera input in- stead of body sensors. PERCLOS [3] measures the percentage of time the eyes remain closed and became widely adopted. Systems using PERCLOS showed high sensitivity to fatigue, especially at night. However, poor lighting, occlusions, and spectacles reduced robustness.

      Facial landmark techniques evolved to detect blinking rate, eye aspect ratio (EAR), and yawning frequency. Park et al. [4] proposed a real-time drowsiness detection model using EAR calculated from 68 facial landmarks. The method effectively detected micro-sleep events but was computationally heavy for low-power devices. Later, deep-learning-based landmark extraction using MediaPipe and Dlib improved speed and accuracy.

      Yawning detection has also been explored. Abtahi et al.

      [5] introduced a mouth aspect ratio (MAR) model to detect prolonged mouth opening. While effective, false positives occurred when drivers talked or sang. Combining multiple facial cues (eye closure + yawning) improved performance, though stable lighting remained necessary.

   3. Driving-Pattern-Based Approaches

      A third category analyzes driving patterns and vehicle behavior. Liang et al. [6] investigated steering wheel move- ment, lane deviation, and pedal pressure as indicators of reduced alertness. These signals were reliable in controlled environments but real-world variability introduced noise. Such methods are most effective when combined with other cues.

   4. Machine Learning and Multi-Modal Approaches

   Machine learning has enhanced detection performance across domains. Ji and Yang [7] developed a hybrid model combining SVM classifiers with facial cues to predict fatigue levels, achieving high accuracy but relying heavily on precise landmark detection. Deep learning, including CNNs trained on datasets like NTHU, UTA-RLDD, and YawDD, improved classification accuracy but required substantial computational power.

   Recent literature highlights multi-modal fusion. Hu et al.

   1. combined eye closure, yawning, head pose, and HRV measurements via decision-level fusion. Hybrid models reduce false positives but require multiple sensors, increasing cost and complexity.

      Modern advancements include deep learning, IoT, and edge computing. Shen et al. [9] proposed a lightweight CNN for embedded systems, achieving efficient performance on devices like Raspberry Pi. Transformer-based architectures and attention mechanisms [10] capture temporal dependencies in driver behavior. Challenges remain in night-time detection, occlusion handling, and balancing cost with accuracy.

      Ref o.	Authors	Methodology	Benefits	imitations
      1	Makeig et al.	EEG-based brainwave analysis	High accuracy; detects early drowsiness	Requires intrusive EEG headgear
      2	Vicente et al.	ECG HRV signal monitoring	Reliable physiological indicator	Needs wearable sensors; motion arti- facts
      3	Govt. PERCLOS Study et al.	Eye-closure percentage detection (PER- CLOS)	Non-invasive; widely validated	Affected by glasses and lighting
      4	Park et al.	EAR using facial landmarks	Real-time; detects micro-sleep	Computationally heavy; occlusion is- sues
      5	Abtahi et al.	MAR yawning-based detection	Detects early fatigue signals	False positives when talking/singing
      6	Liang et al.	Vehicle behaviour (steering, lane devi- ation)	Works without camera	Sensitive to road/weather noise
      7	Ji Yang et al.	SVM + facial feature fusion model	High accuracy hybrid system	Requires clean landmark tracking
      8	Hu et al.	Multi-modal fusion (eye, mouth, HRV)	Reduces false alarms; robust	Expensive setup; complex system
      9	Shen et al.	Lightweight CNN for embedded de- vices	Works on low-power hardware	Lower accuracy than large models
      10	Recent DL Works et al.	Transformer/CNN temporal modelling	Captures long-term patterns	Requires GPU for training
      11	Arefnezhad et al.	EEG + Bayesian filtering	Continuous drowsiness estimation	Intrusive EEG; high computation
      12	Ahmed et al.	Deep Learning visual detection	High accuracy, non-contact	Works poorly in low lighting
      13	Arif et al.	Raw EEG spectral feature extraction	Strong physiological reliability	Requires sensors; simulation data
      14	Rezaee et al.	4-channel EEG vs driving behaviour	Less intrusive EEG setup	Commercial EEG has noise
      15	Florez et al.	CNN eye+mouth on Jetson Nano	Real-time on embedded device	Affected by face occlusion
      16	Albadawi et al.	Survey of DDD methods	Comprehensive technology comparison	No new experimental model
      17	Yu et al.	3D-CNN adaptive driver model	Handles head movement	Camera angle dependency
      18	Salman et al.	Ensemble CNN on YawDD dataset	High performance, robust yawning de- tection	Limited dataset generalization
      19	Bano et al.	EAR + MAR + HOG + SVM	Low computational cost	Less accurate than DL models
      20	Essahraui et al.	Real-time ML facial cue system	Non-intrusive; real-world deployment	Sensitive to occlusion and lighting

      TABLE I

      SUMMARY OF RESEARCH STUDIES ON DRIVER DROWSINESS DETECTION TECHNIQUES

3. METHODOLOGY

   Fig. 1. Flowchart of Driver Drowsiness Detection System

   The proposed system begins with real-time video capture of the driver followed by pre-processing and facial detection. Eye and mouth features are extracted to compute EAR and MAR values for identifying prolonged blinking and yawning patterns. These extracted features are classified to determine

   the driver's alertness level. If the drowsiness score exceeds a predefined threshold, an immediate alert is triggered to prevent potential accidents.

4. FUTURE DIRECTIONS

   1. l. Critical Research Gaps and Limitations

      While driver drowsiness detection has advanced signifi- cantly through multi-modal fusion and deep learning, a syn- thesis of the reviewed literature reveals four persistent gaps that limit real-world reliability and deployment:

      1. Intrusiveness vs. Accuracy Trade-off: The most re- liable detection methodologies, based on physiological signals like EEG [1] and heart rate variability (HRV) [2], are fundamentally intrusive. Their reliance on wearable or contact-based sensors makes them uncomfortable, prone to noise from driver movement [2], and unsuitable for long-term use in everyday commercial vehicles [1].

      2. ack of Robustness to Real-World Conditions: Non- intrusive computer vision systems, while user-friendly, suffer from critical robustness issues. Techniques relying on PERCLOS [3] and facial feature analysis are highly susceptible to poor lighting, partial facial occlusions (spectacles), and external disturbances [3]. Yawning- based detection is vulnerable to false positives due to talking or singing [5], while vehicle-based behavior detection is sensitive to varying road conditions [6].

      3. Computational Burden for Edge Deployment: Ad- vanced ML approaches such as high-precision landmark extraction [4], SVM classifiers [7], and CNN models

         1. impose substantial computational demands. These requirements make real-time deployment challenging on low-power embedded systems [4], [7], [9].

      4. Complexity of Hybrid Systems: Although multi-modal fusion significantly improves accuracy and reduces false alarms [8], the requirement for multiple heterogeneous sensors increases system complexity, installation diffi- culty, and overall cost [8].

      5. imited Generalization Across Drivers and Environ- ments: Systems trained on controlled laboratory datasets struggle to generalize across diverse drivers, fatigue patterns, and lighting conditions [11], [13]. Bayesian EEG-based approaches [11] demonstrate promise but lack large-scale real-road validation.

      6. Sensitivity to ighting and Face Orientation in Deep Models: Deep-learning visual systems still experience major performance drops when the driver's face is partially visible or lighting is poor [12], [17]. Current methods fail to ensure reliability during nighttime driv- ing.

      7. Dataset imitations and Real-World Validation Gap: Many studies use limited datasets such as YawDD [18], affecting the robustness of trained models. Embedded CNN implementations [15], [18] remain insufficiently tested under real-world highway or long-duration driving environments.

      8. Hardware Constraints in ow-Power Systems: Ap- proaches designed for edge platforms like Jetson Nano demonstrate feasibility [15] but suffer from processing bottlenecks, limiting real-time multi-modal inference [9], [15].

      9. High Cost and Integration Barriers in Multi-Sensor Fusion: Systems combining cameras, physiological sen- sors, and steering behavior [8], [20] result in improved accuracy but introduce increased cost, wiring complex- ity, and installation effort.

      10. ack of Benchmark Standardization: Research works use different datasets, thresholds, and evaluation criteria [16], making cross-comparison difficult and slowing the adoption of standardized industry-level systems.

   2. Future Research Directions

      Based on the identified limitations, the next phase of re- search should strategically focus on overcoming the practical barriers to widespread, reliable driver drowsiness detection.

5. FUTURE RESEARCH DIRECTIONS

   Although the current system successfully leverages Medi- aPipe facial landmarks to compute Eye Aspect Ratio (EAR) and Mouth Aspect Ratio (MAR) with a real-time alarm mechanism, several enhancements can further improve robust- ness, accuracy, and usability. The following directions outline potential improvements for future work:

   1. Integration of ightweight Deep earning Models: Secondary verification using a lightweight CNN or trans- former can reduce false positives in occluded or noisy frames.

   2. Dynamic EAR/MAR Threshold Personalization: Adaptive thresholds for each driver can be learned based on their baseline blinking and yawning patterns.

   3. ight-Time and ow- ight Enhancements: Incorpo- rating infrared (IR) camera support or low-light enhance- ment algorithms can improve nighttime reliability.

   4. Improved Robustness Against Occlusion: Multi-view or occlusion-resistant tracking methods can handle masks, sunglasses, or hands-over-face.

   5. Temporal Deep earning for Micro-Sleep Detection: LSTM, GRU, or transformer-based networks can model temporal patterns in EAR/MAR sequences for better micro-sleep detection.

   6. Integration with Driving Behavior Metrics: Fusion with steering, lane deviation, and braking data can enhance detection accuracy.

   7. Context-Aware Decision Fusion: Using road context such as traffic, time of day, and journey duration to adapt alert thresholds intelligently.

   8. Multi-Modal Alert Mechanisms: Combining audio alarms with seat vibrations, dashboard indicators, or mobile notifications to improve response effectiveness.

   9. Stress and Emotion Differentiation: Differentiating drowsiness from strong emotions like sadness or stress using facial expression models.

   10. Energy-Efficient Edge Deployment: Optimizing the system for devices like Jetson Nano or Raspberry Pi using quantization and model pruning.

   11. arge Real-World Dataset Collection: Gathering di- verse driving videos under various lighting conditions, demographics, and vehicle types for better generaliza- tion.

   12. Privacy-Preserving Detection: Performing on-device inference and encrypted storage to protect driver identity.

   13. Fail-Safe Redundant Safety Mechanisms: Designing mechanisms to safely slow or stop the vehicle or alert emergency contacts if the driver does not respond.

   14. Driver Workload and Task Complexity Modeling: Incorporating cognitive workload measures to reduce false positives in multitasking situations.

   15. Human-Machine Interaction Based Alert Optimiza- tion: Personalizing alerts (sound, vibration) based on driver sensitivity, urgency, and situational parameters.

   16. Explainable AI for Trust: Using attention maps or visual explanations so users or manufacturers understand why an alert was triggered.

   17. Uncertainty Estimation: Introducing probabilistic models, e.g., Bayesian deep learning, to provide con- fidence scores for detections.

   18. Hybrid Cloud-Edge Frameworks: Balancing onboard inference with cloud analytics for continuous improve- ment.

   19. Standardization and Regulatory Compliance: Defin- ing testing protocols, safety validation, and compliance for automotive deployment.

   20. User Adaptation and Personalization Over Time:

   Allowing the system to adjust to a driver's changing habits or conditions via continual learning.

6. CONCLUSION

This review paper has comprehensively charted the evolu- tion of Driver Drowsiness Detection (DDD) methodologies over the past two decades, highlighting a pivotal shift from intrusive physiological sensing to non-contact computer vision and advanced deep learning approaches.

Our synthesis of the literature across the three primary domains-physiological signals, facial features, and driving behavior-confirms several key findings. Physiological mea- surements such as EEG [1], [11], [13], [14] and HRV [2] provide the most accurate correlation with fatigue but rely on intrusive sensors, which limits practical deployment [1], [2], [13]. Non-intrusive visual cues, including PERCLOS [3], EAR [4], [19], and MAR [5], [19], show promising user acceptance but remain sensitive to lighting conditions, occlusions, and distinguishing true fatigue from natural driver behavior [5], [7], [12], [15].

Hybrid and multi-modal approaches that combine facial features, physiological indicators, and vehicle behavior [6]-

1. have demonstrated higher accuracy and robustness. How- ever, these systems often involve complex sensor setups, synchronization requirements, and high computational costs, which can hinder real-world adoption [4], [7]-[10], [15], [17]. Lightweight and embedded-friendly architectures [9], [15], [17] have been explored to reduce computational over- head while maintaining reasonable accuracy, yet balancing performance, efficiency, and robustness remains an ongoing challenge [10], [12], [16], [18].

Additionally, ensemble models and adaptive frameworks [18], [20] show potential in improving generalization across diverse drivers and driving conditions, while survey stud- ies [16] emphasize the need for standardization, large-scale datasets, and real-world testing. Systems capable of distin- guishing micro-sleeps from short eye closures [4], [12], [19] and incorporating probabilistic confidence measures [11], [17] are critical for safe deployment.

In conclusion, the successful transition of DDD from labora- tory research to widespread commercial integration depends on addressing these practical trade-offs. Future efforts must focus on: (i) developing ultra-lightweight, robust models [9], [15], [17], (ii) creating adaptive and generalized multi-modal fusion techniques [6]-[8], [18], [20], and (iii) ensuring deployment- ready frameworks that balance accuracy, efficiency, and cost [10], [12], [16], [19]. Such strategies will enable highly accurate, cost-effective, and user-friendly driver monitoring systems capable of real-time fatigue detection under diverse driving conditions.

1. Government PERCLOS Study et al., "Eye closure percentage detection (PERCLOS) for driver alertness," 2008.

2. Park et al., "EAR using facial landmarks for micro-sleep detection,"

   Sensors, 2016.

3. Abtahi et al., "Yawning detection using MAR for driver fatigue," lEEE Transactions on Affective Computing, 2013.

4. Liang et al., "Vehicle behavior analysis for non-visual drowsiness detection," Transportation Research, 2015.

5. Ji Yang et al., "SVM + facial feature fusion for drowsiness," Pattern Recognition Letters, 2014.

6. Hu et al., "Multi-modal fusion for driver fatigue detection," lEEE Transactions on lntelligent Transportation Systems, 2016.

7. Shen et al., "Lightweight CNN for embedded drowsiness detection,"

   Neural Computing and Applications, 2019.

8. Recent DL Works et al., "Transformer/CNN temporal modeling for driver fatigue," 2020.

9. Arefnezhad et al., "EEG + Bayesian filtering for continuous drowsiness estimation," 2018.

10. Ahmed et al., "Deep learning visual detection of driver drowsiness,"

    lEEE Access, 2020.

11. Arif et al., "Raw EEG spectral feature extraction for driver fatigue," 2017.

12. Rezaee et al., "4-channel EEG vs driving behavior analysis," 2016.

13. Florez et al., "CNN eye+mouth detection on Jetson Nano," 2019.

14. Albadawi et al., "Survey of drowsiness detection methods," 2020.

15. Yu et al., "3D-CNN adaptive driver model," 2018.

16. Salman et al., "Ensemble CNN on YawDD dataset," 2019.

17. Bano et al., "EAR + MAR + HOG + SVM for low-cost drowsiness detection," 2017.

18. Essahraui et al., "Real-time ML facial cue system," 2018.

REFERENCES

1. E. Makeig et al., "EEG-based brainwave analysis for drowsiness detec- tion," Journal of Neuroscience Methods, 2000.

2. J. Vicente et al., "ECG and HRV monitoring for driver fatigue," lEEE Transactions on Biomedical Engineering, 2014.