An Integrated Deep Learning-Based CNN–LSTM Framework for Driver Monitoring and Automotive Safety System

The Intelligent transportation systems require automotive safety solutions that can accurately detect hazards and respond in real time under diverse traffic and environmental conditions. This study presents an integrated deep learning‑based framework combining Convectional Neural Networks (CNNs) and Long Short‑Term Memory (LSTM) networks for driver monitoring and automotive safety. In the proposed architecture, a CNN module extracts spatial features from dual camera inputs driver‑facing and road‑facing views to automatically identify vehicles, pedestrians, lane markings, and road anomalies without handcrafted feature engineering. These frame‑level representations are then processed by an LSTM module that models temporal dependencies across consecutive observations, enabling early recognition of forward collision risk, lane departure, and abnormal driving behavior. To ensure robust road‑scene representation, the system is trained and validated on more than 1,000 road and driver images covering varied urban and highway scenarios, lighting conditions, and traffic densities.The system is evaluated using large‑scale automotive datasets and a YOLOv8‑based detection backbone. Experimental results show that the driver monitoring component achieves 98.21% accuracy, 97.84% precision, 97.36% recall, and a 97.60% F1‑score, indicating reliable classification of driver state. The object detection pipeline attains 96.92% mAP@0.5, 97.55% precision, and 96.81% recall, while lane detection and collision warning modules reach 98.43% and 97.68% accuracy, respectively. Distance estimation remains within ±0.18 m, with an overall processing speed of 42 frames per second and an inference time of 23.8 ms per frame, demonstrating suitability for real‑time deployment. These results suggest that the hybrid CNN–LSTM framework provides superior detection accuracy, early warning capability, and robustness compared with standalone CNN and conventional machine learning approaches.