應用於駕駛安全監控之人臉表情辨識系統於雙核心嵌入式平台上之實現

Abstract

隨著經濟的逐步起飛，人民生活的水準也不斷提高，汽車作為交通工具日益普及，成為人們日常生活中不可或缺的一部分，然而車輛的普遍性與便利性，也使得交通堵塞的情況更為嚴重，進而導致由憤怒情緒引起的交通事故逐年增加，而「怒路族」一詞也隨之產生，因此，如何有效的辨識駕駛人的情緒，並且分析及監控具有危險性的駕駛行為，為本研究所要解決的。 本論文主要分成三大部分：截取人臉特徵、人臉表情辨識、駕駛危險係數推估。第一部分，利用興趣區域之區域三元化圖型(Area of Interest Local Ternary Patterns)對貢獻度高的人臉區塊，抓出其紋理特徵，之後利用此資訊；在第二部分以線性支持向量機為辨識核心來進行表情辨識；第三部分利用車外行車行為、車內駕駛表情做為資訊，透過模糊推論系統推估出危險係數。本研究實作於Devkit8500D雙核心嵌入式平台上，並且於室內、戶外與車上的場景中均可辨識成功，表情辨識演算法可達到97%以上的偵測率。

With the development of our society and economy, cars as important transportation become normal and essential in people’s daily life. However, the universality and convenience of vehicles make traffic jams more serious, and traffic accidents caused by angry emotion increase year by year. The word “Road Rage” already comes to light. So, how to identify the driving emotion, analyze and monitor the dangerous driving behavior needs to be solved which the kernel of this study. In this thesis, a facial expression recognition algorithm by Area of Interest Local Ternary Patterns (AI-LTP) and Support Vector Machine (SVM) is presented. Our system could be mainly separated into three parts: extraction, recognition and driving danger-level estimate. In extraction part, AI-LTP is applied to extract the face’s appearance features by high contribution of face block. In recognition part, Support Vector Machine is used to distinguish the seven different facial expressions: neutral, anger, disgust, fear, happiness, sadness, surprise. The third part uses vehicle driving behavior and driving expression as input information. Then estimate the danger-level through fuzzy inference system. The proposed system is successfully implemented on the Devkit8500D embedded platform, and fully tested in indoor, outdoor, and driving environments. The experimental results show the accuracy ratio of 97% for the face expression recognition.