Real-Time Drowsiness Detection Using Smartwatch Sensor Data and Machine Learning

Driver drowsiness is a leading and highly preventable cause of road accidents worldwide. This thesis therefore asks: Can smartwatch data accurately classify four distinct drowsiness levels with at least 90% accuracy in realistic driving conditions? Can we use this information to detect and prevent driver drowsiness? To investigate this topic, we analyze a dataset composed of 169,466 samples recorded with commercial smartwatch sensors. Each measurement combines Heart Rate (HR), Heart Rate Variability (HRV), and Respiration Rate (RR) with a synchronised arousal label that is discretised, via equal-width binning, into four classes spanning from full alertness to pronounced drowsiness. Our principal methodological innovation is a lightweight real-time processing system that streams incoming data through a sliding window, performs real-time feature scaling, and feeds the resulting features directly to the classifier. This arrangement eliminates the need for batch processing and allows continuous inference on resource constrained embedded hardware. Two supervised learners are proposed: a Random Forest (RF) ensemble tuned for rich, non-linear interactions, and a radial basis function Support Vector Machine (SVM) configured for soft margin, multiclass separation. Testing on individual drivers showed a mean accuracy of 88 % (max 99 %) for the RF, markedly outperforming the SVM at 77 % (max 94 %); both models maintained good performance across all drowsiness levels. While the RF falls just shy of the 90 % target on average, its peak accuracy and the consistent margin over the SVM highlight the promise of ensemble techniques for physiological state recognition. These results demonstrate that low-cost wearable devices combined with lightweight machine learning algorithms achieve accuracy suitable for real-world use. By signaling transitions from alertness to early drowsiness within seconds, the proposed system can be integrated into driver assistance systems, fleet safety dashboards, or consumer smartwatch applications, thereby reducing fatigue-related crash risk and associated public health costs.