Implementation of a driver drowsiness prediction system. The Prediction Pipeline has two stages. In stage one we detect the drivers face, using a pre-trained retinaface model, the detected face is than fed into a ResMaskNet model, which was pretrained on the FER2013 dataset, and finetuned for Driver Drowsiness Prediction. We fine-tune the model on a custom dataset called FL3D (dataset consisting of frames from NITYMED, hand annotated by us). For more information please read Frame Level Driver Drowsiness Prediction.
To evaluate models ability to generalize to out of distribution data, we used the following dataset: roboflow-driver-drowsiness-detection.
You can take a look at the rest of predictions, including failure cases in: Predicitons.
The following is the performance on the ood dataset:
(face-detector) mean_iou: 0.51, n_detector_fails: 2
(driver-state predictor) acc: 0.83, precision: 0.83, recall: 0.76, f1-score: 0.78