Automated Edge-to-Cloud Retraining of MobileNetV2 for Real-Time Truck Driver Drowsiness Detection

Truck drivers who are experiencing drowsiness can cause fatal accidents. There are 459,248truck drivers involved in accidents between January 2022 and October 2026 in Indonesia. Degradation of performance due to environmental variations commonly occurs in driver monitoring systems. This paper presents an automated Edge-to-Cloud retraining framework that is designed for real-time driver drowsiness detection. The system combines local edge inference on a Raspberry Pi 5 with a cloud-based MLOps retraining engine on Google Colab Pro. Face landmark tracking is performed using the MediaPipe Face Landmarker.Binary classification relies on a MobileNetV2 architecture. The initial model was trained on the public Yawning Detection Dataset (YawDD). Local edge calculates the Mouth Aspect Ratio (MAR) and evaluates MobileNetV2 prediction confidence.A yawn is detected when prediction probability exceeds 0.9 and MAR reaches or exceeds 0.6 across five consecutive frames. Validated events trigger an asynchronous WhatsApp alert to driver monitoring and upload high-confidence image frames to Google Drive. Google Colab Pro performs a two-stage retraining pipeline with dynamic class weighting when accumulated field images reach five frames per class. The trained model is updated and deployed only when the test Macro F1-Score is higher thanthe previous Macro F1-Score. After conducting five retraining cycles, the peak test Macro F1-Score improved from a baseline of 93.12% to 94.09%. Raspberry Pi 5 performance averaged 5.82 FPS (174.24 ms), 44.00% CPU utilization, and 795.00 MB RAM consumption, and took approximately 0.87 seconds to make a decision.