Publication: Vision-based driver drowsiness detection sytem using YOLOV5
Loading...
Date
2024-07
Authors
Nur Izzati binti Jailani
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Ensuring driver safety on the roads is important, with drowsiness posing a significant threat. Fatigue can compromise a driver's focus, reaction time, and decision-making abilities, elevating the risk of accidents. To tackle this issue, this project aims to build vision-based drowsiness detection for monitoring driver drowsiness. In this system, the task of monitoring is automated by using YOLOv5 model object detection algorithm to detect drowsiness under two different lighting conditions. This technology can monitor the driver's facial expressions, eye movements, and other behavioural cues, providing timely alerts and interventions to prevent accidents caused by drowsy driving. The development of the system can be divided into five stages. The first stage of development begins with collecting driver face images in both drowsy and awake conditions for datasets. After that, second stage continue with the pre-processing collected image datasets in previous phase by image resize and image labelling. Then, third stage continue with training custom YOLOv5 model by using the image datasets pre-processed in previous phase. Next, fourth stage continue with validate and test the trained YOLOv5 model on Google Colab. At this stage, if the accuracy of YOLO model is low and not as expected, the first stage of collecting driver face images is restarted to add more images to training datasets until the accuracy meet expectation. Lastly, the development ends with the fifth stage, which involves evaluating the performance of YOLOv5 through both quantitative and qualitative analysis. The YOLO model demonstrated excellent performance in terms of detection accuracy under both good and poor lighting conditions, consistently maintaining reliable confidence scores. When tested on video, the model exhibited excellent results with a detection accuracy rate of 95.62%, an F1 score of 94.06%, precision of 95%, and recall of 93.14%. Additionally, the false detection rate was very low, with a false positive rate of 3.33% and a false negative rate of 6.86%, both well below 10%.