Real-time Driver’s Eyes Tracking System using Semantics-based Vague Image Representation and Fully Connected Neural Network on Single Chip

• Main Authors: Ting, YI-SIANG, 丁弈翔
• Other Authors: YU, YING-HAO
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/4c3zjh

碩士 === 國立中正大學 === 電機工程研究所 === 107 === Abstract For a smart vehicle system design, driver’s attention monitoring system is essential to advanced driver assistance system (ADAS). Such technology can be achieved by using face tracking to detect distraction of driver. The computing process involves complicated feature extraction and pattern recognition so that design concepts of small dimension, high computing performance, and low power consumption are required in order to be implement in a vehicle. For this, this study focuses on the way to realize an efficient driving eyes tracking algorithm on a single chip. In this research, the algorithm of Semantics-based Vague Image Representation(SVIR) with a Fully Connection Neural network(FCN) are implemented on a single Field Programmable Gate Array(FPGA) chip to track driver’s eye direction in real-time with economical hardware resources usage. The reliability for eyes direction detection can be verified by a high separation from FCN. Furthermore, the proposed system also avoids the use of DSP with low logic elements and memory usages. Therefore, it is especially suitable for small embedded systems. According to the experimental results of this study, with 640*480 image resolution and 80 image frames per second, it only consumes 0.52 us to finish eyes tracking. In addition, high reliability can be found from 20% separation to the other eyes directions with 70% similarity threshold at the output of FCN. These saliences demonstrate the feasibility for advanced driver assistance systems in the future. Keywords: Human Eyes Monitoring, Semantics-based Vague Image Representation, Fully Connected Neural Network, Real-time Processing System, Advanced Driver Assistance System