Driver state monitoring is an important component of intelligent transportation systems because changes in alertness and visual attention can affect driving performance and safety. This study aimed to design and experimentally evaluate a vision-based, multimodal driver-state monitoring system integrating temporal analysis and edge artificial intelligence for real-time driver-state classification and laboratory safety-response support. The experimental dataset comprised 375 driving videos collected from 48 licensed drivers aged 22–56 years, yielding 18,250 valid frames. Three driver states were considered: alert, drowsy, and distracted. Two independent evaluators performed the state annotation, with substantial inter-rater agreement (Cohen’s kappa = 0.91). To reduce information leakage, data were partitioned at the driver level, ensuring that observations from the same driver did not occur across the training, validation, and test sets. Visual and complementary features included eye aspect ratio, blink rate, eye-closure duration, mouth aspect ratio, head-pose angles, gaze direction, physiological information, and vehicle-dynamic variables. Temporal information was incorporated using a 15-frame window corresponding to 0.5 s at 30 frames/s. The evaluated model achieved an accuracy of 97.84%, precision of 97.21%, sensitivity of 97.62%, specificity of 98.14%, F1-score of 97.41%, Matthews correlation coefficient of 0.956, Cohen’s kappa of 0.963, and ROC-AUC of 0.988. On the edge-computing platform, the mean frame-processing time was 31.4 ms, while the measured end-to-end latency was 116 ms. The laboratory safety-response experiment produced a response time of 0.92 s. These findings indicate that the proposed framework can provide accurate driver-state classification and low-latency processing under the tested experimental conditions. Further research should evaluate the system with larger and more diverse driver populations, independent environments, multimodal ablation experiments, uncertainty estimation, and field-based safety validation. Further research should evaluate the system with larger and more diverse driver populations, independent environments, multimodal ablation experiments, uncertainty estimation, and field-based safety validation.