Driver drowsiness is a leading cause of traffic accidents worldwide, highlighting the urgent need for accurate and real-time detection systems to enhance road safety. Recent advances in deep learning have shown great promise, particularly with attention-augmented convolutional neural networks (CNNs). However, most existing studies focus on a single attention mechanism in isolation, providing limited insight into their comparative effectiveness. In this work, we present a systematic and comprehensive framework for integrating, analyzing, and benchmarking multiple attention mechanisms within a custom CNN architecture tailored for driver drowsiness detection. Specifically, we investigate four representative modules: Squeeze-and-Excitation (SE) channel attention, Spatial Attention, Efficient Channel Attention (ECA), and the hybrid Convolutional Block Attention Module (CBAM). A baseline CNN without attention is implemented to enable direct assessment of the added value of each mechanism. Furthermore, we benchmark our framework against six widely adopted transfer learning models, including VGG19, DenseNet169, ResNet50V2, InceptionV3, MobileNet, and InceptionResNetV2. Extensive experiments conducted on the NTHU-DDD2 dataset under real-world driving conditions demonstrate that attention mechanisms significantly enhance classification performance. Among them, CBAM achieved the best trade-off between accuracy and efficiency, reaching 99.63% accuracy, an AUC of 0.9999, and competitive inference latency suitable for real-time deployment. ECA also performed strongly, validating the role of lightweight attention in improving recall and training stability. By providing the first comprehensive comparative study of diverse attention mechanisms in this domain, this work establishes a robust benchmark and demonstrates that hybrid attention-augmented CNNs, particularly CBAM, are highly effective for driver drowsiness detection. These findings advance the development of intelligent and safety-critical transportation systems, while outlining pathways for future research toward robust, efficient, and deployable solutions.
Hassan et al. (Thu,) studied this question.