Driver Drowsiness Detection Using Fixed and Dynamic Thresholding

Globally, drowsiness is a major cause of road accidents. Although extensive work has been performed in this field, this paper enhances the accuracy of drowsiness detection for long-distance journeys using the Dlib-based facial feature detection algorithm. The Dlib shape predictors extract the facial features that help in calculating essential parameters, i.e., Eye Aspect Ratio (EAR) and Mouth Opening Ratio (MOR). Two proposed algorithms; fixed thresholding and dynamic frame thresholding; are proposed. The two methods use both MOR and EAR parameters to decide the driver’s drowsiness level. The fixed threshold approach alerts the driver of drowsiness once the EAR and/or MOR values meet a certain criterion. The threshold for EAR is set to 0.15, while MOR is set to 0.4. A driver is notified as drowsy when: (1) EAR is less than 0.15 (closed eyes), (2) MOR is greater than 0.4 (yawning), (3) EAR is less than 0.15, and (4) MOR is greater than 0.4 (closed eyes and yawning simultaneously). The accuracy and sensitivity in detection using fixed thresholding were 89.4 and 96.5, respectively, using 1000 images. Dynamic frame thresholding algorithm involves a counter to track a set of several consecutive frames that meet the criterion before sending a warning. Further, the consecutive number of frames is adjusted as the time elapses to increase detection accuracy and better communication of a drowsy state. Four (4) videos, each of 30 minutes, are passed through the Dynamic frame thresholding algorithm to evaluate its response. The accuracy and sensitivity of drowsiness detection are (93.4) and (89), respectively, using 686 images.