This paper will present a Hybrid CNN Transformer model which is a spatial feature of the face after which temporal dynamics of the behavioral aspects are combined to provide a robust real-time drowsiness detection. Using the Public datasets including YawDD and Kaggle driver drowsiness, the model exploits the Eye Aspect Ratio (EAR) and facial landmarks to recognize the indicators of drowsiness among them protruded closeness, the frequency of blinking, and the slightest muscle alterations. The CNN part is aimed at capturing discriminative spatial features on eye and face regions, and the Transformer part is dedicated to capturing long-range temporal relations across frame sequences, which is a dynamical progression of drowsiness. The preprocessing steps, such as facial landmark normalization, adaptive histogram equalization, and temporal signal smoothing, are aimed at improving the generalization and making the block more noise tolerant. Also, there is fusion of the EAR and mouth aspect data with deep features at the middle level so as to enhance stability of prediction. The experimental scores are verified to perform higher performance than baseline CNN and LSTM models and have high accuracy, precision, and F1-scores on varying lighting and head-pose conditions.