Driver drowsiness is one of the major factors that lead to serious traffic accidents, developing a reliable and real-time detection system is very important. In this paper, we propose a three-step vision-based pipeline for detecting driver drowsiness by combining computer vision and deep learning methods. First, an adaptive keyframe extraction approach based on histogram analysis and chi-square distance is applied to automatically select the most informative frames from a video sequence. Then, MediaPipe Face Mesh is used to extract facial landmarks, and key geometric features such as Eye Aspect Ratio (EAR), Mouth Aspect Ratio (MAR), and blink-related attributes are computed to represent drowsiness-related facial behavior. Finally, a hybrid CNN–Transformer model is designed to capture both short-term and long-term temporal patterns from the extracted feature sequences. The proposed model effectively distinguishes between Alert and Drowsy states, achieving 97.5% accuracy, a high F1-score, and a low false alarm rate, outperforming several traditional machine learning and deep learning baselines. The main contribution of this work is the integration of adaptive keyframe selection with an attention-based hybrid deep learning architecture, which makes the system efficient and suitable for real-time implementation in intelligent vehicle safety systems.