201140511 六、發明說明： 【發明所屬之技術領域】 本發明係關於-種_測方法，特別為一種先偵測出 人臉位置，再找出人臉區域中瞳孔中心位置，操取出眼睛影 像及眼睛特徵參數，以這些眼睛特徵參數訓練一支持向量機 謂分類器’用來分類與判斷眼睛是睁眼或是閉眼，最後 計算·L0S，若PERCL0S超過一個預先設定門播值時， 則認為在打瞌睡 【先前技術】 影像式瞌睡偵測是近幾年來新興起的一項研究技術， 主要應用在判斷駕駛是否㈣睡’其非接觸性及高方便性的 優勢，廣泛的受到研究學者與產業界的高度重視，期盼能於 駕車安全上有優異的表現。—般的影像式暖睡制系統由 •影像操取模組同步的操取駕駛臉部正面影像，分析二連續影 像藉由5十算水平位移顯著之圖素的水平和垂直投影決定駕 駛者人臉區域；由眼睛與人臉面部之-般人體測量比率估算 眼目月區域，计异眼睛區域的水平和垂直投影決定眼皮位置， 由眼皮位置δ十算駕駛者連續眨眼間隔，若眨眼間隔超過設定 1檻值f启文動—警告提醒駕驶者。這種做法已揭露於專利 持續皿控車輕駕駛者之方法及裝置，以偵測防止駕驶者打 瞌睡」(專利公開號：436436)中。另外，如論文「An 一oved 201140511201140511 VI. Description of the Invention: [Technical Field] The present invention relates to a method for detecting a face, in particular, for detecting a position of a face first, and then finding a position of a pupil center in a face region, and taking out an eye image and Eye feature parameters, training a support vector machine with these eye feature parameters, the classifier is used to classify and judge whether the eye is blinking or closed eyes, and finally calculate L0S. If PERCL0S exceeds a preset gatecast value, it is considered Dozing off [Prior Art] Image-style sleepiness detection is a research technology emerging in recent years. It is mainly used to judge whether driving is (4) sleepy. Its non-contact and high convenience advantages are widely studied by researchers and industry. The community attaches great importance to it and hopes to have excellent performance in driving safety. The image-based warm sleep system is controlled by the image manipulation module to capture the frontal image of the driver's face. The analysis of the two consecutive images determines the driver's person by the horizontal and vertical projection of the significant horizontal displacement of the pixels. Face area; the eye-moon area is estimated by the ratio of the eye to the face-like body measurement. The horizontal and vertical projections of the eye area determine the eyelid position. The eyelid position δ is calculated as the driver's continuous blink interval. If the eye gap interval exceeds the setting 1 f value f start the text - warning to remind the driver. This practice has been disclosed in the patented method and apparatus for controlling the light driver of a vehicle to detect that the driver is prevented from falling asleep (Patent Publication No. 436436). In addition, as the paper "An oved 201140511
Real Time Eye State Identification System in Driver Drowsiness Detection」（參考文獻1)中，由影像擷取模 組同步的擷取駕駛臉部正面影像，利用連接式Adab〇〇st 方法偵測人臉位置，計算人臉影像的水平投影決定眼睛位 置，將眼睛影像以一門檻值二值化後，計算邊緣複雜度決定 是睁眼或閉眼，如此可計算PERCL0S，若PERCL〇s超過一預 先設定的門檻值則認為在打瞌睡以上這些影像式的瞌睡偵 測方法與系統都有一個共通點，計算人臉區域影像的水平投 影決定眼睛位置，當人臉有些微的傾斜或戴粗框眼鏡時，會 叙重影響人臉區域影像的水平投影，進而無法精準的偵測眼 睛位置’眼睛位置谓測不準自然無法正確偵測駕駛瞌睡。 由此可見，上述習用方式仍有諸多缺失，實非一良善 之設計’而亟待加以改良。 本案發明人鑑於上述習用方式所衍生的各項缺點，乃 函思加以改良創新，並經多年苦心孤諸潛心研究後，終於成 功研發完成本件瞌睡偵測方法。 準確性 本發明之目的即在於提供一種瞌睡偵測方法其中利 用輕射對稱轉換方㈣測瞳孔中心位置，可改善當前的目鱼睡 偵測方法中利用人臉影像之水平投影偵測眼睛位置的缺 點’用以提升眼睛位置偵測之準確性，進而提升_偵測的 201140511 達成上述發明目的之瞌睡偵，丨 i悄，則方法’係利用影像擷取 模組同步的擷取駕駛臉部正 此曲衫像，利用連接式Adab〇〇st 方法（參考文獻2)偵測人臉位置表 职仙直運接式Adaboost方法的 運作方式為利用預先訓練好的Α η。l 不于的Adaboost模型判斷影像中 不同位置與大小的正方形框是否為人臉。八_。⑽模型的 詞練方式為先以人工收集人臉和非人臉影像，操取影像的特 徵參數’人臉㈣時所採用的特徵參數為（參 考文獻3,此為小波轉換係數的—種)，擷取特徵參數後利 用Adaboost演算法挑選出有用的特徵參數，Adab〇〇st演 算法是兩類別（人臉與非人臉）的分類演算法，其原理為從眾 夕的人臉特徵參數（Haar Feature)中以迭代更新的方式一 人-人的挑選有用的特徵參數，而一個特徵參數搭配一個門 檻值就能視為一個弱分類器，每次挑選有用的特徵參數其實 就疋挑選有用的弱分類器，最後由弱分類器以權重組合的方 式成為一個強分類器，此即為Adaboost人臉判斷模型。 而連接式Adaboost的原理是再結合多個Adab〇〇st人臉 判斷模型達到以更少且有用的人臉特徵參數快速過渡影像 中非人臉的正方形框，其假設是認為影像中非人臉的正方形 框占絕大多數，在實際應用時此方法的確能達到加速的效 果0 4貞測出人臉位置後利用輻射對稱轉換方法（參考文獻4 ) 偵測瞳孔中心位置，以提升眼睛位置偵測準確性，進而達到 201140511 提升瞌睡偵測準確性之目的。輻射對稱轉換方法的原理為偵 測人臉影像中輻射形狀的中心點。其方式為先計算人臉影像 中所有位置的灰階值變化方向（梯度方向），預先定義瞳孔在 人臉中的可能半徑大小，從人臉影像中判斷每個位置是否為 其他位置所指的梯度方向的中心，若是，則代表此位置报有 可能是曈孔中心，以此方式在人臉影像中可能會偵測出好幾 個輻射形狀的中心點，例如曈孔或鼻孔，由人臉五官相對位 置可篩選出瞳孔正確位置，此方法不會受到人臉些微傾斜或 戴粗框眼鏡的影響。 瞳孔中心位置獲得後，以此位置擷取出眼睛矩形區域 影像，利用（LBP)Local Binary Pattern方法（參考文獻 5)擷取眼睛區域影像特徵參數，LBp方法的原理為統計眼 睛影像中每個位置的灰階值與其㈣8個位置的灰階值 之大小關係’計算方式為以眼睛影像中每個位置為中心比 較其和8個相鄰位置的灰階值大小，並將結果以數字i 或0表示，因此比較結果是一個長度為8的二位元字 串，將其轉成十進制其範圍介⑨〇心5，最後以直方圖統 計每個數字出現的次數就是眼睛影像的特徵參數。 擷取出眼睛影像的特徵參數後利用預先訓練好的支持 向量機（SVM)模型（參考讀6)判斷眼晴狀態是睁眼或閉 眼。SVM帛型的訓練方式為先以人卫收料眼和閉眼的影 201140511In the Real Time Eye State Identification System in Driver Drowsiness Detection (Reference 1), the image capture module is used to capture the front image of the driver's face, and the connected Adab〇〇st method is used to detect the face position and calculate the person. The horizontal projection of the face image determines the position of the eye. After the eye image is binarized by a threshold, the edge complexity is determined to be blinking or closing the eye. This can calculate PERCL0S. If PERCL〇s exceeds a preset threshold, it is considered In the above-mentioned sleepy detection methods and systems, there is a common point. The horizontal projection of the image of the face region determines the position of the eye. When the face is slightly tilted or the glasses are worn, the effect will be emphasized. The horizontal projection of the image of the face area makes it impossible to accurately detect the position of the eye. 'The position of the eye is too accurate to detect the driving drows properly. It can be seen that there are still many shortcomings in the above-mentioned methods of use, which is not a good design and needs to be improved. In view of the shortcomings derived from the above-mentioned conventional methods, the inventors of the present invention have improved and innovated, and after years of painstaking research, they finally succeeded in researching and developing this method of detecting sleepiness. Accuracy The purpose of the present invention is to provide a doze detection method in which the center position of the pupil is measured by using a light-weight symmetry conversion side (4), which can improve the current eye-sleep detection method and use the horizontal projection of the face image to detect the eye position. Disadvantages 'to improve the accuracy of eye position detection, and then improve the detection of 201140511 to achieve the above-mentioned purpose of the dormant detection, 丨i quiet, the method 'use the image capture module to synchronize the driving face is positive This type of shirt uses the connected Adab〇〇st method (Reference 2) to detect the face position table. The operation of the Adaboost method is to use the pre-trained Αη. l The Adaboost model does not determine whether the square frame of different positions and sizes in the image is a human face. Eight_. (10) The word training method of the model is to collect the face and non-face images manually, and the characteristic parameters used when the image's feature parameter 'human face (4) is taken (Ref. 3, which is the wavelet transform coefficient) After taking the feature parameters, the Adaboost algorithm is used to select useful feature parameters. The Adab〇〇st algorithm is a classification algorithm of two categories (face and non-face). The principle is the face feature parameters from the public eve ( Haar Feature) is a one-person selection of useful feature parameters in an iterative update manner, and a feature parameter can be regarded as a weak classifier with a threshold value. Each time a useful feature parameter is selected, it is useful to select a weak one. The classifier is finally a weak classifier by weight combination to form a strong classifier. This is the Adaboost face judgment model. The principle of connected Adaboost is to combine multiple Adab〇〇st face judgment models to achieve a fast transition of non-human square frames in the image with fewer and useful facial feature parameters, which is assumed to be non-human faces in the image. The square frame is the majority, and in practice, this method can achieve the acceleration effect. 4 4 After detecting the face position, the radiation symmetry conversion method (Reference 4) is used to detect the pupil center position to enhance the eye position detection. Accuracy, and then achieve 201140511 to improve the accuracy of doze detection. The principle of the radiation symmetry conversion method is to detect the center point of the radiation shape in the face image. The method is to first calculate the direction of the grayscale value change (gradient direction) of all the positions in the face image, and predetermine the possible radius of the pupil in the face, and determine whether each position is pointed out by other positions from the face image. The center of the gradient direction, if it is, indicates that the position may be the center of the pupil. In this way, the center point of several radiation shapes may be detected in the face image, such as the pupil or the nostril. The relative position can be used to screen out the correct position of the pupil. This method is not affected by slight tilting of the face or wearing thick-rimmed glasses. After obtaining the position of the pupil center, the image of the rectangular region of the eye is taken out at this position, and the image feature of the eye region is extracted by the (LBP) Local Binary Pattern method (Reference 5). The principle of the LBp method is to count each position in the eye image. The relationship between the grayscale value and its (four) grayscale value of 8 positions is calculated by comparing the grayscale value of the eight adjacent positions centered on each position in the eye image, and the result is represented by the number i or 0. Therefore, the result of the comparison is a two-bit string of length 8, which is converted into decimal and its range is 9 cents. Finally, the number of occurrences of each digit in the histogram is the characteristic parameter of the eye image.撷 After taking out the feature parameters of the eye image, use the pre-trained Support Vector Machine (SVM) model (Ref. 6) to determine whether the eye condition is blinking or closing the eye. The training method of SVM type is to first receive the eye of the eye and close the eyes of the eye. 201140511
像’利用LBP方法掏取出特徵參數後利用支持向量機（SVM) 演算法做兩類別(睁眼和閉眼)的訓練得至"VM帛型。SVM 演算法是兩類別的分類演算法，其原理是找出能區分兩類別 :徵參數資料的高維空間平面，且兩類別的特徵參數資料中 靠近高維空間平面的資料和此高維空間平面的距離達到最 遠0 經由影像操取模組操取連續的人臉影像後，透過上述 流程可判斷出每張眼睛影像是睁眼或閉眼，最後計算 PERCLOS，# PERGLGS料—預先設定的門錄則認為在 打θ盍睡。 【實施方式】 請參閱圖-所示，係本發明瞌睡偵測方法模組圖包 含影像操取1组1G'人臉制模組n、輻射對稱轉換眼睛 谓測模組12、眼睛特徵擷取模组13、支持向量機（svm)眼 睛狀態分類模組14、PERCL0S計算模組i 5。 請參閱圖二所示，係本發明瞌睡偵測方法實施流程 圖。透過影像擷取模組10拍攝駕駛人臉影像，利用人臉 偵測模組11進行人臉位置定位得到人臉區域影像，這裡 的人臉偵測是採用連接式Adaboost方法；利用輻射對稱 轉換眼目月偵測模組12在人臉區域影像中進行曈孔中心位 201140511 置定位’曈孔中心位置獲得後，以曈孔中心位置操取出眼睛 區域影像’再利用眼睛特徵擷取模組13 f十算眼睛區域= 像的特徵參數，這裡的眼睛特徵參數之擷取係採用 Binary Pattern方法；眼睛特徵參數經由支持向量機（sv心 眼睛狀態分類模組14判斷眼睛狀態是睁眼或閉眼，支持 向量機（SVM)眼睛狀態分類模組14係由支持向量機（svm) 演算法對眼睛資料庫影像做兩類別的訓練所得到之眼睛狀 態分類模組，而這裡的兩類別指的是睁眼及閉眼的影像。 經由影像擷取模組1 〇擷取的連續人臉影像，透過上 述的流程後每張人臉影像都能判斷其為睁眼或閉眼如此經 由PERCL0S計算模組15計算PERCL〇s，pERCL〇s計算模 組1 5會計算在一段時間内閉眼的秒數（例如十秒），並設定 一門檻值（例如八秒），若PERCL〇s超過一預先設定的門檻 值，則認為是打瞌睡，反之則是未打瞌睡。 本發明所提供之瞌睡偵測方法，與其他習用技術相互比 較時’更具備下列優點： 1.本發明可提升眼睛偵測的適應性，使其在駕駛頭部 些微傾斜或戴粗框眼鏡時仍能正確偵測駕駛瞳孔中心位 置。 2.本發明可提升瞌睡偵測準確性，使其應用於瞌睡偵 測系統更可確保駕駛行車安全。 201140511 上列詳細說明乃針對本發明之一可行實施例進行具體 §兄明’惟該實施例並非用以限制本發明之專利範圍，凡未 脫離本發明技藝精神所為之等效實施或變更，均應包含於 本案之專利範圍中。 综上所述’本案不僅於技術思想上確屬創新，並具備習 用之傳統方法所不及之上述多項功效，已充分符合新穎性 及進步性之法定發明專利要件，爰依法提出申請，懇請貴 局核准本件發明專利申請案，以勵發明，至感德便。 參考文獻 l.Tianyi Hong, Huabiao Qin and Qianshu Sun, An Improved Real Time Eye State Identification System in Driver Drowsiness Detection, IEEE International Conference On Digital ObjectAfter using the LBP method to extract the feature parameters, the support vector machine (SVM) algorithm is used to perform the training of the two categories (blinking and closing the eyes) to the "VM type. The SVM algorithm is a two-category classification algorithm. The principle is to find a high-dimensional spatial plane that can distinguish two categories: the parameter data, and the data of the two types of characteristic parameter data close to the high-dimensional space plane and the high-dimensional space. The distance of the plane reaches the farthest distance. After the continuous face image is processed by the image manipulation module, the above process can be used to determine whether each eye image is blinking or closed, and finally calculate PERCLOS, #PERGLGS material-pre-set door The record is considered to be sleeping at θ. [Embodiment] Please refer to the figure-show, the module diagram of the doze detection method of the present invention includes image manipulation 1 set of 1G' face system module n, radiation symmetric conversion eye presence module 12, eye feature capture Module 13, support vector machine (svm) eye state classification module 14, PERCL0S calculation module i 5. Please refer to FIG. 2, which is a flow chart of the implementation of the doze detection method of the present invention. The image capturing module 10 is used to capture the driver's face image, and the face detection module 11 is used to position the face to obtain the face region image. The face detection is performed by the connected Adaboost method; The monthly detection module 12 performs the pupil center position in the face area image 201140511. Positioning 'After the pupil center position is obtained, the eye area image is taken out at the pupil center position'. The eye feature extraction module 13 f is used. Calculate the eye area = the characteristic parameter of the image, where the eye feature parameter is captured by the Binary Pattern method; the eye feature parameter is determined by the support vector machine (the sv heart-eye state classification module 14 determines that the eye state is blinking or closed eyes, support vector The machine (SVM) eye state classification module 14 is an eye state classification module obtained by the support vector machine (svm) algorithm for performing two types of training on the eye database image, and the two categories here refer to the blinking and Closed-eye image. Through the image capture module 1 captures the continuous face image, through the above process, each face image can be judged as 睁Or closing the eye to calculate PERCL〇s via the PERCL0S calculation module 15, the pERCL〇s calculation module 15 calculates the number of seconds to close the eye for a period of time (for example, ten seconds), and sets a threshold (for example, eight seconds), if If PERCL〇s exceeds a preset threshold, it is considered to be dozing off, and vice versa is not dozing. The sleep detection method provided by the present invention, when compared with other conventional techniques, has the following advantages: 1. The invention can improve the adaptability of the eye detection, and can correctly detect the center position of the driving pupil when the driver's head is slightly tilted or wears the thick-frame glasses. 2. The invention can improve the accuracy of the sleep detection and apply it to The sleep detection system is more secure in driving. 201140511 The above detailed description is directed to a possible embodiment of the present invention, which is not intended to limit the scope of the invention. The equivalent implementation or change of the spirit of the art should be included in the scope of the patent in this case. In summary, the case is not only innovative in terms of technical thinking, but also has a habit. The above-mentioned multiple functions that are not in line with the traditional methods have fully met the statutory invention patent requirements of novelty and progressiveness. If you apply in accordance with the law, you are requested to approve the application for the invention patent to encourage the invention and to the sense of convenience. l.Tianyi Hong, Huabiao Qin and Qianshu Sun, An Improved Real Time Eye State Identification System in Driver Drowsiness Detection, IEEE International Conference On Digital Object
Identifier, pp. 1449-1453, 2007. 2. Paul Viola, Michael Jones, Robust Real-Time Face Detetion, International Journal of Computer Vision, 2004. 3. C.Papageorgiou, M. Oren, and T. Poggio. A general framework for object detection. In International Conference on Computer vision, 1998. 9 201140511 4. Gareth Loy and Alexander Zelinsky, Fast Radial Symmetry for Detecting Points of Interest, IEEE Transactions on Pattern Analysis and Machine Intelligence, VOL. 25, NO. 8, August 2003. 5. Ojala, T., Pietikainen, M., Harwood, D. : A Comparative Study of Texture Measures with Classification Based on Feature Distributions, Pattern Recognition 29(1996) 51-59. 6. J.C. Burges, A Tutorial on Support Vector Machines for Pattern Recognition, Data Mining and Knowledge Discovery, vol. 2(2), 1998, pp 121-167. 【圖式簡單說明】 請參閱有關本發明之詳細說明及其附圖’將可進一步瞭 解本發明之技術内容及其目的功效；有關附圖為： 圖一為本發明瞌睡偵測方法模組圖。 圖二為本發明瞌睡偵測方法實施流程圖。 【主要元件符號說明】 1 〇影像摘取模組 11人臉彳貞測模組 12輻射對稱轉換眼睛偵測模組 13眼睛特徵擷取模組 14支持向量機(謂)眼睛狀態分類模組 201140511 15 PERCL0S計算模組Identifier, pp. 1449-1453, 2007. 2. Paul Viola, Michael Jones, Robust Real-Time Face Detetion, International Journal of Computer Vision, 2004. 3. C. Papageorgiou, M. Oren, and T. Poggio. A general Framework for object detection. In International Conference on Computer vision, 1998. 9 201140511 4. Gareth Loy and Alexander Zelinsky, Fast Radial Symmetry for Detecting Points of Interest, IEEE Transactions on Pattern Analysis and Machine Intelligence, VOL. 25, NO. August 2003. 5. Ojala, T., Pietikainen, M., Harwood, D. : A Comparative Study of Texture Measures with Classification Based on Feature Distributions, Pattern Recognition 29 (1996) 51-59. 6. JC Burges, A Tutorial On Support Vector Machines for Pattern Recognition, Data Mining and Knowledge Discovery, vol. 2(2), 1998, pp 121-167. [Simplified Description of the Drawings] Please refer to the detailed description of the present invention and its accompanying drawings' The technical content of the present invention and the purpose of the present invention are as follows: FIG. 1 is a sleep detection method of the present invention. Photos. FIG. 2 is a flowchart of implementing the doze detection method of the present invention. [Main component symbol description] 1 〇Image extraction module 11 Human face detection module 12 Radiation symmetry conversion Eye detection module 13 Eye feature capture module 14 Support vector machine (referred to) Eye state classification module 201140511 15 PERCL0S calculation module
nn