Hough Transform based Drowsy Driver Detection System and Embedded System Implementation

• Main Authors: Chien-Wei Chang, 張建煒
• Other Authors: Chih-Peng Fan
• Format: Others
• Language: zh-TW
• Published: 2014
• Online Access: http://ndltd.ncl.edu.tw/handle/32202515033604868919

碩士 === 國立中興大學 === 電機工程學系所 === 102 === In this thesis, we propose a drowsy detection system which is unaffected by light. In any kind of weathers, even if drivers wear glasses or sunglasses, the system can still detects whether a driver is drowsy or not. By the camera to capture facial images, the system will start an alarm when the driver is drowsy after a long-time driving, and the proposed system reminds the driver to get rest for safety. To make sure that the system works all day long, we use a near infrared ray camera as an input device. Instead of RGB scale images, we only use gray scale images as input data to prevent side effects by light. The proposed system is separated into two subsystems: the drowsy detection and the nod detection. The drowsy detection uses eye status as the input information. If the driver closes his/her eyes for a period of time, the system will start an alarm. The nod detection uses head motions as the input information when the driver wears sunglasses to make us misjudge the eye information. If one of the subsystems detects that the driver is drowsy, the system will start an alarm. To improve the proposed drowsy detection system, a lot of software optimization techniques are used, which include the down sampling for facial region searches, the integral images to speed-up searching of Haar-like features, the improved SIO algorithm, using the look up table to replace repeating tasks of Hough transform, the fast threshold decision for open/closed eye status, and the two-dimension to one –dimension convolution. In our experiments, a personal computer with 3.4GHz Quad-core CPUs and 4GB memories is used for simulations, and the processing speed is up to 63 frames per second. Eight different video sequences in the car circumstance are used for tests, including normal and drowsy cases, daytime and nighttime cases, and glasses and non-glasses cases. The detection accuracy rate of the open/closed eye status is up to 86%, and the detection accuracy rate of the awake/drowsy status is up to 91%. Finally, the drowsy detection system is implemented on the Zedboard embedded platform. We use the GUI on Linux system for real-time image processing and recognition. By the NIR camera as the input device and the HDMI screen as the output device, the processing speed is up to 16 frames per second after software optimizations.