A Comparative Analysis of Object Detection Models for Drowsiness Detection: Evaluating Raw, Processed, and Hybrid Image Datasets with Faster R-CNN

In Mexico, last year, 5.10% of accidents attributed to human factors were related to drowsiness. Consequently, this study investigates the detection of objects to identify visible signs of drowsiness, such as the eyes and mouth, employing a Faster R-CNN model with ResNet50 and FPN. A comparative analysis uses three datasets: raw images, processed images, and a hybrid set comprising both. The results indicate that although the model trained on raw images exhibited the lowest loss and superior performance for certain classes, the model trained on processed images demonstrated the highest efficiency in training time. Meanwhile, the model trained on the hybrid set achieved a balanced performance, integrating the strengths of both preceding models.