Embedded Eye-State Classification and Fleet Event Logging for a Driver Drowsiness Alert Prototype

Abstract

Camera-based drowsiness alerting requires reliable visual measurements and an operational process for event recording. This study developed a preliminary Raspberry Pi prototype that combined a compact MobileNetV2 eye-state classifier, consecutive-frame alert logic, and fleet event dashboard. A balanced subset of 20,000 MRL eye images was allocated at the image level to the training, validation, and test partitions in a 75/15/10 ratio. The classifier used 64 × 64 grayscale inputs, a width multiplier of 0.35, two-stage transfer learning, and TensorFlow Lite deployment. The compact classifier contained approximately 492,000 parameters and was trained for 23 epochs. On 2,000 test eye images, it correctly classified 1,957 images, achieving 97.85% accuracy, 98.40% closed-eye sensitivity, 97.30% open-eye specificity, and a receiver operating characteristic area under the curve of 0.9977. A Raspberry Pi 5 desk-based session processed 2,824 frames in 202.4 seconds, with a reported average of 13.95 frames per second and 1.24 ms per eye-classification call. The FastAPI and SQLite platform supported the event history, evidence images, and company, vehicle, and device management. Fifteen event records were observed, of which three lacked resolved vehicle associations. These findings support the preliminary technical feasibility of the reported tests. This study did not establish subject-independent performance, drowsiness diagnosis, or operation in a moving vehicle.