A HYBRID EDGE-AI FRAMEWORK FOR LOW-COST DRIVER DROWSINESS DETECTION USING EAR AND MOBILENETV2 IN NIGERIAN DRIVING ENVIRONMENTS

Driver drowsiness is a significant contributor to road accidents, posing a severe challenge to public safety in Nigeria. Existing drowsiness detection systems are heavily reliant on non-African data, often expensive and inflexible to driving environments and conditions in developing countries, which makes them unsuitable for widespread local adoption in Nigeria and other African nations. This research addressed this gap by implementing and evaluating a cost-effective drowsiness detection using locally gathered real-world driver images from Nigeria. The dataset was further classified into awake and drowsy based on the eyes behavioral pattern and percentage of eye closure. The system utilizes a low-cost webcam to monitor the driver states, processing the video stream with Python, OpenCV, and Dlib. Drowsiness was detected using a robust hybrid methodology: computationally efficient geometric metrics, specifically the Eye Aspect Ratio (EAR) for prolonged eye closure was used for initial detection. These outcomes are then cross-validated by a trained lightweight CNN architecture (MobileNetV2) to enhance accuracy and reduce false positives. The trained model of the developed system achieved an overall accuracy of 97.0% with a precision of 97.0%, a recall rate of 97.0%, and an F1-score of 97.0% on a randomly split testing dataset. The system was tested and implemented in a real-time driving environment, confirming its ability to reliably detect drowsiness events while maintaining a manageable false alarm rate, as evidenced by only 177 misclassifications of 8167 test instances. Based on the findings, the high performance indicates that the accessible, low-cost hybrid approach is an effective countermeasure in a resource-constrained environment that can greatly improve safety on Nigerian roads.