DESIGN OF FACE-BASED DROWSY DRIVER DETECTION SYSTEM WITH COMPUTER VISION AND MACHINE LEARNING

• Main Author: Deni Adhitama, Yulian
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/46339
• Institution: Institut Teknologi Bandung

At present, a warning system to detect signs of sleepy drivers is developing throughout the world and there are some cars that are equipped with technology that can improve user safety, especially analyzing the driver's condition and giving a warning when the driver shows signs of drowsiness. However, the system is built-in and not all types of cars have the system. So that drivers who want to improve driving safety cannot buy the system alone, but must buy a new car that is equipped with the system. Therefore, a system is developed that can analyze the condition of the driver and provide a warning when the driver shows drowsy signs that can be installed as an additional feature in the car, namely the In-Car Assistive Technology for Drowsy Driver (AssystDrive) system. One of the subsystems in AssystDrive is the driver's face detection subsystem. This subsystem analyzes the condition of the driver based on the driver's face image and classifies it into 3 conditions namely awake, low vigilance, and drowsy. Awake conditions state that the driver is awake. When the driver is in this condition, no warning is given. Low vigilance conditions state that the driver is not concentrating. When the driver is in this condition, the path detection subsystem and the object detection subsystem will be active. The drowsy condition states that the driver is drowsy. When this condition is active, the system will give a warning to the driver. Taking the driver's face image is done in realtime using a camera placed on the front dashboard of the car. Infrared filters on the camera are removed and infrared LEDs are added to the camera so that the camera can take pictures in dark conditions. Images taken by the camera are processed using Odroid XU4. Alert notifications are given through the buzzer sound and displayed on an LCD screen placed on the car's dashboard. Classification of driver conditions is made using machine learning algorithms and computer vision. Testing is done in realtime and by using video input. Realtime test results show that images can be taken and processed in bright and dark light conditions with a maximum computing delay of 500 milliseconds. Tests with video input were done with 115 videos which were labeled awake and low vigilance and obtained an accuracy of around 70%. The sound generated by the buzzer when the system gives a warning is around 55dB.