Drowsiness Detection Using Yolov5 Algorithm

Yeong, Weng Liang (2022) Drowsiness Detection Using Yolov5 Algorithm. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Yeong Weng Liang.pdf Restricted to Registered users only Download (4MB)

Abstract

Drowsiness driving is one of the major causes of road accidents and the best way to detect drowsiness is through Behavioural Based Measurements (BBM). In BBM, the facial characteristics of the driver is analysed through number of yawning and rate of eye blinking. Drowsiness detection based on facial analysis is subject to variety of factors such as low illumination level and object hindrance wore by individuals that could potentially affect the accuracy of detection. Drowsiness detection aimed to be applied in real life requires high accuracy and detection speed. In this project, a drowsiness detection system using deep learning technique that could achieve high accuracy and speed are aimed to be achieved. The algorithm chosen in this project is YOLOv5 algorithm which is a powerful and fast algorithm that is written in Python programming language. Due to YOLOv5 algorithm is being relatively new, the proposed method will also be reviewed and evaluated. YOLOv5 was trained and tested using custom dataset which comprised of images from Roboflow dataset and YawDD dataset. Few techniques such as Cosine Annealing, Warm-Up strategy, Complete Intersection Over Union (CIoU) and transfer learning are applied in this project. The evaluated model are YOLOv5s which achieves mAP50 of 92.9% and inference time of 17.4ms and YOLOv5m which achieves mAP50 of 95.0% and inference time of 45.3ms. Hence, YOLOv5s is a more suitable algorithm with hardware constraint and prioritizing speed.