Universitas Dian Nuswantoro (UDINUS), through researchers Giap, Yo Ceng; Muljono; Affandy; Basuki, Ruri Suko; Azies, Harun Al; Isnanto, R. Rizal; and Dewi, Deshinta Arrova, examined an attention-guided lightweight convolutional neural network design based on MobileNetV2 for real-time driver drowsiness detection. The activity addressed the need for accurate and rapid detection on Edge-IoT platforms with limited resources. Driver drowsiness was described as a major cause of traffic accidents.
The study compared a Squeeze-and-Excitation-enhanced MobileNetV2 with a baseline model and a structurally optimized version. The optimized version used Depthwise Separable Convolution, Bottleneck blocks, and Expansion layers. Experiments were conducted on 500 images. The results indicated that channel attention enhanced feature discrimination, while structural optimization produced the most resilient trade-off between accuracy and latency. Statistical validation used 95% confidence intervals and two-proportion Z-tests to substantiate the significance of the enhancements.
The proposed models supported real-time inference despite their small size, described as about 2.6 million parameters and 315 million FLOPs. The findings suggested that structural optimization was more important than attention mechanisms when designing lightweight convolutional neural networks for embedded driver monitoring. This UDINUS research activity therefore addresses driver drowsiness detection and the design of lightweight models for resource-limited Edge-IoT systems. The source does not provide further information on implementation context, institutional outputs, or documented follow-up.