請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/35938
標題: | 電腦視覺眼部偵測為基礎之駕駛疲勞警示系統 Computer Vision-based Eye Detection and Warning System for Driver Fatigue |
作者: | Yi-Ru Chen 陳怡如 |
指導教授: | 張堂賢 |
關鍵字: | 駕駛疲勞,眼睛偵測,倒傳遞類神經網路,模糊隸屬函數,Telematics客服中心, Driver Fatigue,Eye Detection,Backpropagation Neural Network,Fuzzy Membership Function,Telematics Call Center, |
出版年 : | 2005 |
學位: | 碩士 |
摘要: | 本研究針對駕駛疲勞建構一套以電腦視覺為基礎的眼部偵測警示系統,利用紅外線攝影機紀錄駕駛人眼睛的運動特性。當解讀出有疲倦駕駛行為時,車內警示系統會適時發出警訊促使駕駛人恢復警覺;如果駕駛人仍無法保持清醒狀況,系統會透過通訊設備自動發送緊急訊息至Telematics服務中心,客服人員在接獲通報後將主動與車端聯繫,藉由與駕駛人攀談的方式以消除其駕駛疲勞。本系統將依序構建「眼部偵測模組」與「疲勞警示模組」。
「眼部偵測模組」中,本研究構建一套監控駕駛人警覺狀況的影像處理程序以逐步確認眼睛位置,包含:臉部偵測、眼睛偵測以及眼睛追蹤。首先,臉部偵測的目的為減少畫面中的搜尋範圍,其演算法在光線充足下是利用橢圓形樣板比對出最似人臉區域;光線不足下改以紅外線輔助光源投射於駕駛頭部並利用動態門檻值將人臉與背景分離。待臉部偵測完畢,初步眼睛偵測是利用累積影像相減法逐一篩選出候選點,再進一步以邊界偵測器過濾出最具眼睛特徵的目標物而決定眼睛位置完成標記。最後利用α-β濾波預測下一次目標物將出現的位置以獲得更佳的偵測結果,此追蹤模式使得模組更為穩健。 在「疲勞警示模組」中,閉眼延時百分率(PERCLOS)與平均眨眼速率(AECS)為判斷駕駛警覺程度的重要參數,本研究利用倒傳遞類神經網路的學習訓練功能獲得適性門檻值以解決因人而異的眼部運動特性。警覺程度與行車速率將共同決定警示程度,利用模糊隸屬函數產生的警示門檻值可避免發出不必要的警訊干擾駕駛人,譬如在車行速度緩慢下,駕駛人的警覺程度可稍有懈怠而不會產生警示。 本研究共測試四個環境,其中三個案例的眼部偵測成功率達97.8%;另一個案例則因駕駛人佩戴眼鏡造成失誤率攀高,眼鏡上的光流會造成誤判,但其成功率也達84.8%。故本系統在大部分的環境下具可行性。 This study developed computer vision-based eye detection and warning system. The purpose of this system is to perform detection of driver fatigue. By mounting a small infrared camera on the dashboard, a driver’s eyelid movements are recorded. While this driver is tired and no longer in condition to drive, a warning signal would be issued, and at urgent time, an emergency call would deliver to Telematics Call Center automatically where agents could offer a real-time help to link the driver by voice-activated cellular phone. Eye Detection Module contains three major parts: face detection, eye detection, and eye tracking. Face detection algorithms are developed to reduce searching areas. These methods are based on both ellipse-template matching in daytime and dynamic threshold selection in nighttime. After determining the face, an accumulative different image method is used to extract eye candidates in primary step. Further, an edge detector is used to find the most possible eye targets to separate other objects such as eyebrows. Finally, to make the module more robust, a α-β filter is used to track eye targets in subsequence frame. This eye detection module is used to calculate the breadth of eye open to determine vigilance levels of a driver. To monitor a driver fatigue, both Percentage of Eye Closure over Time (PERCLOS) and Average Eye Closure Speed (AECS) are calculated. Because eyelid movement characters are distinct from people, a Backpropagation Neural Network is applied to learn and train the adaptive vigilance levels by different drivers. Combing vigilance levels and vehicle speeds, a warning threshold is determined by Fuzzy Membership Functions, which can prevent unnecessary warning signals. For example, although a driver is very tired, the alert wouldn’t work if the car moves slowly. There are total 4 test cases in the experimental results. Only one case failed because this driver wearing glasses which cause optical-flow to reduce eye detection accuracy. But in this failed case, the probability of success is still around 84.8 %. The probabilities of success are all above 97.8 % in the other cases. It shows this system is feasible to most cases. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/35938 |
全文授權: | 有償授權 |
顯示於系所單位: | 土木工程學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-94-1.pdf 目前未授權公開取用 | 2.76 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。