Table 2 The summary of the analyzed related works.
Ref | Year | Dataset | Classification model | Accuracy | Strengths and Weaknesses |
|---|---|---|---|---|---|
2019 | Self-prepared dataset | RF and non-linear SVM | RF: 88.37% to 91.18% | Strengths: Good performance with varying epoch lengths. Weaknesses: SVM accuracy is consistently lower | |
2019 | CelebA, YawDD | Multiple CNN-kernelized correlation filters method | 92% | Strengths: High accuracy in a variety of conditions, robust to environmental variations. Weaknesses: Limited to CNN models, lacks broader evaluation across other architectures | |
2020 | 300-W dataset | Mamdani fuzzy inference system | 95.5% | Strengths: Incorporates fuzzy logic for drowsiness detection, useful for real-time applications. Weaknesses: May struggle with fine-tuning or handling complex image data without enhancement | |
2020 | Self-prepared thermal | SVM and KNN | SVM: 90% | Strengths: Non-invasive method, useful for thermal monitoring in different lighting conditions. Weaknesses: Thermal imaging may require high-end equipment, and KNN struggles in complex environments. | |
image dataset | KNN:83% | ||||
2020 | Self-prepared ZJU dataset | FD-NN, TL-VGG16, and TL-VGG19 | FD-NN: 98.15%, TL-VGG16: 95.45%, TL-VGG19: 95% | Strengths: Impressive performance for fatigue detection, able to capture fine-grain eye movement features. Weaknesses: Limited dataset, reliance on predefined classifiers | |
2020 | Self-prepared dataset (DROZY database) | Multilayer perceptron, RF, and SVM | SVM: 94.9% | Strengths: Good feature extraction using basic neural networks for fatigue detection. Weaknesses: Lack of deep learning-based approaches, might miss subtle facial cues. | |
2021 | NTHU-DDD video dataset | Deep-CNN-based ensemble | 85% | Strengths: Ensemble learning improves detection, high accuracy in diverse environments. Weaknesses: Lower overall performance compared to transformer-based models | |
2022 | CEW, ZJU, MRL | Dual CNN Ensemble (DCNNE) | CEW: 97.56%, ZJU: 97.99%, MRL: 98.98% | Strengths: High performance across multiple datasets with ensemble methods. Weaknesses: May not generalize well to datasets outside of the tested range | |
2022 | UTA-RLDD dataset | RNN and CNN | 60% | Strengths: Low computational cost, suitable for mobile applications. Weaknesses: Low accuracy, particularly for complex real-time applications | |
2022 | Self-prepared dataset for traffic signs | CNN | 98.53% | Strengths: Strong accuracy in traffic-related scenarios, applicable to many monitoring systems. Weaknesses: Lacks focus on drowsiness detection, limited to specific contexts | |
2022 | NTHU-DDD | CNN + LSTM | 97.3% | Strengths: Efficient for sequential data analysis, effective in dynamic environments. Weaknesses: Struggles with long-duration analysis or sustained predictions in real-time | |
2022 | Public dataset using gas sensor, temperature sensor, and digital camera | t-SNE for feature extraction + Isolation Forest (iF) for anomaly detection | 95% | Strengths: Effective detection using only normal (non-drunk) data, Handles nonlinear, high-dimensional data well, Unsupervised approach suitable for real-time detection. Weaknesses: t-SNE is computationally intensive and not ideal for real-time processing, Model performance may vary with different sensor quality or configurations | |
2023 | Drowsiness dataset | CNN and VGG16 | CNN: 97%, VGG16: 94% | Strengths: High performance in detecting drowsiness from real-time data. Weaknesses: Limited to VGG16 architecture, potential underperformance in new data types | |
2023 | MRL | VGG16, VGG19, and 4D | VGG16: 95.93%, VGG19: 95.03%, 4D: 97.53% | Strengths: Strong results across multiple configurations, robust for real-time driver drowsiness detection. Weaknesses: Dependence on specific VGG-based models may limit flexibility in dynamic environments | |
2023 | NTHUDDD dataset | RF, SVM, and sequential NN | RF: 99%, SVM: 80%, 4D: 96% | Strengths: RF offers excellent performance, especially for simple fatigue detection scenarios. Weaknesses: SVM underperformed significantly, less robust across diverse environmental conditions | |
2024 | Public dataset using gas sensor, temperature sensor, and digital camera | ICA for feature extraction + Kantorovitch Distance (KD) + DEWMA for anomaly detection; XGBoost for SHAP analysis | F1-score = 98% | Strengths: Does not require labeled data (semi-supervised) High sensitivity using DEWMA with nonparametric threshold, SHAP adds explainability to the model, Effective on non-Gaussian multivariate data. Weaknesses: Complexity due to integration of multiple techniques, DEWMA and KD may require careful tuning for different datasets, Potentially computationally intensive for real-time systems | |
2024 | NTHU-DDD | VGG19 | 96.51% | Strengths: Efficient in various lighting and environmental conditions. Weaknesses: Performance variation across datasets, still limited by fixed network architectures | |
2024 | YawDD, MRL | VGG16 and CNN | VGG16: 95.85%, CNN: 96.54% | Strengths: High performance in real-time detection with various feature extraction methods. Weaknesses: Not ideal for low-complexity devices, might need more robust processing power | |
2024 | MRL | CNN, InceptionV3, and MobileNetV2 | CNN: 96%, MobileNetV2: 97%, InceptionV3: 98% | Strengths: Excellent performance with quick response times, particularly in driver monitoring. Weaknesses: InceptionV3 and MobileNetV2 still face computational trade-offs |
