M364920 五、新型說明： 【新型所屬之技術領域】 ，尤指一種紅外線 本創作與人臉識別技術有關 照明的立體人臉識別系統。 【先前技術】 ^ μ #臉識別系統，在安全監控系統中有著非常廣 托丄…'用。目前的人臉識別技術對於應用環境存在 闽ΛΑ ϋ Ϊ制，使得這種技術报難在實際應用中大範 ，，推廣。其中’最大的限制在於不能適應環境的 巨^變化。例如，環境光線、人臉姿態及表情的不 同II化，造成傳統技術方法的識別率高，幾乎 具備實用價值。 、請參閱圖1，為習知技術為使用單一攝影機的人 臉識別裝置。其透過單一的攝影機13擷取使用者人 臉’透過傳輸設備1 6將影像傳至監控中心1 7 ,分析 厂，人臉影像内的特徵值與資料庫之人臉樣板影像 進行人臉辨識。然而，實際應用環境的複雜性，尤 其在非理想環境光照條件下時’大多數人臉識別系 統必然遇到識別性能下降問題。隨著光線的變化， ❿同一張人臉的資訊可能產生很大的變化，往往造成 人臉特徵的差別。這也是限制現有人臉識別系統真 正可以供使用者實際使用的最重要的原因。大部分 ，研究工作在於對現有可見光人臉識別系統進行改 進，以減輕環境光照的影響。雖然取得了一定的進 〜步’但成效不佳。加上，以單張人臉影像虛擬的3D 人臉模型·缺乏人臉深度資訊。 而之所以目前常見作法，即以單一攝影機擷取 使用者人臉’取得了單張人臉影像後再加以模擬成 三維人臉模型，之後再以此人臉模型為準去辨識一 待測人臉影像，而此習用技術誤判率甚高的原因在 3 M364920 於，實際應用的環境光的複雜性，尤其在非 境光照條件下時，意即光源的數量、光源的 度、光源的顏色等條件，與當初建立資剩· $ 攝狀況、環境有所不同時，而這些的不同， 照射角度或是數量的不同，必定使人臉上的 化，意即陰影的位置、明暗位置與建檔資料 同，而習用技術以單一攝影機拍攝單張的影 維影像’當使用這個光影與建檔者不同的二 去模擬成三維影像，再去與資料庫内原本的 像對比’勢必會得到待識別人並非本人的結 言之，當有人冒用該本人的資料欲進行識^ 有可月b因為ί衣境光的不同’反而使得待測二 可以被模擬成一個與建檔的三維影像相近的 而造成該冒充者被誤判為本人’這在安全上 嚴重的問題’由此可見，習用技術所造成大 臉識別系統遇到識別性能下降問題，是十 的’這也是限制現有人臉識別系統真正應用 要的，因。大部分研究工作在於對現有可見 ^ ^系統進行改進，以減輕環境光照的影響 取付一定的進步，但成效不佳。 詳έ之’另一個技術上的問題在於，人 Μ 據人臉上的特徵，由於人臉變化複雜 二I 2和擷取上十分困難。這諸多因素使得 臉Η別Γ項極富挑戰性的課題。前述習用技 研1糸統多是針對二維照片或動態視訊序 別ί影像處理技術為基礎；但是，二維 別的影變ίϊ”礙，無法解決上述的問 本身3二炝產生這些問題的主要原因是人臉 面π ί 2 Γ的，而二維的照片是對三維曲面 又心、、、告果’投影的問題即在於缺乏影像 理想環 照射角 時的拍 不論是 光影變 有所不 像是二 維影像 二維影 論； 時， 維影 結果 是十 多數 分危 的最 光人 胎 4 M364920 中會失去一部分 方法最大的區別 述’例如人臉的 構等等。通過更 過程中的誤判率 光照無關性和姿 人臉的基本特性 情的影響，從而 此可知，三維人 目前在這一領域 資訊，在此過程 -維識別與傳統的 訊可以更好的描 '及各點之間的結 較好的解決識別 維人臉模型具備 能夠正確反映出 二維結構不受表 '•臉特徵表述。由 可以很好的解決 鲁瓶頸。 重要資訊。採用三 就在於，人臉的資 特徵點的深度資訊 全面的資訊，可以 問題，同時由於三 態無關性的特點， ，同時人臉主要的 形成相對穩定的人 臉模型的識別方法 慣用的二維圖像的 為=解決人臉姿態及表情的不同變化，有人提 ^1用^維雷射掃描器擷取模塑物體的三維座標並 f其=維模型；或是以結構光攝影機，在物體上 生固定樣式的條紋，再依據條紋的形狀計算物體 位置與建構三維模型；或是利用人臉影像序 =二維，模型。其中，有一種建構三維臉部的 技巧=利用單張或多張的人臉影像來建構被模塑人 ，的三維臉部模型。由預先準備的通用三維人臉模 ^，根據從二維影像以互動、半自動或自動擷取到 的人臉特徵點的位置來調整通用三維人臉模型的形 狀。依據人臉影像中各器官特徵點來重建三維模型 •的重要特徵點，將一般三維人臉模型的三角化頂點 間的關係，加上二維人臉影像的紋路及貼圖技術 、建立個人的三維人臉模型。然而，以單張人臉影 虛擬的三維人臉模型·缺乏人臉深度資訊。 此外’光源是影響人臉識別的一項重要因素。 上^的描述已經說明了隨著光線的變化，同一張臉 的資訊可能產生很大的變化，往往造成特徵的差 另】使彳于本人識別成非本人，而非本人卻識別成本 5 M364920 人’對安全造成極大的威脅。 此外，若以上述的三維雷射掃描器擷取三維人 η。^然此設備系統受環境光照影塑較工维但 ::成ΐί和計算複雜度很高，雷射光對人 i剎田舳h 煬 不犯滿足實際系統的需 要利用…、，工外線或遠紅外線影像的方法，易受環 Κί傻念和健康狀態的影響，使得獲得的 的變化’在實際的應用系統中性 月b亚+好。 r格a丨大的人臉辨識在人臉的角度上有著 用文單茫二傻的次^二維人臉模型表示的做法大都採 擬的人臉系統以供多角度的辨嘈ώ '、虛 依f 一張正面影像去做模擬，欠缺深度的資訊， ^ 擬的三維人臉時，當觀看的角度改ΐ 比:面視角時)’就可發現人臉影像長 就會下滑。-般最常見的情 ίίΐΐ: 歪斜、或鼻子的形狀高度有麫 誤’而如此重要的五官之一的誤 Α 、-曰 !J i的錯誤，而這裡出現誤判的原因就ΐ因為缺ί 人臉深度資訊。 y疋口馬缺乏M364920 V. New description: [New technical field], especially a kind of three-dimensional face recognition system related to illumination and creation of face recognition technology. [Prior Art] ^ μ #Face recognition system, which is very widely used in security monitoring systems... The current face recognition technology has a defamatory system for the application environment, making this technology difficult to be widely used in practical applications. Among them, the biggest limitation is that it cannot adapt to the huge changes in the environment. For example, the different II of ambient light, face pose and expression make the recognition rate of traditional technical methods high and almost practical. Referring to Fig. 1, a conventional method is a face recognition device using a single camera. It captures the user's face through a single camera 13 and transmits the image to the monitoring center through the transmission device 16 to analyze the feature values in the face image and the face image of the database for face recognition. However, the complexity of the actual application environment, especially in non-ideal ambient lighting conditions, most of the face recognition systems must encounter the problem of reduced recognition performance. As the light changes, the information on the same face may change greatly, often resulting in differences in facial features. This is also the most important reason to limit the existing face recognition system that can actually be used by the user. For the most part, research efforts have been made to improve existing visible light face recognition systems to mitigate the effects of ambient light. Although it has achieved a certain degree of progress, it has not achieved good results. In addition, the virtual 3D face model with a single face image lacks face depth information. The reason why the current common practice is to capture the user's face with a single camera' has acquired a single face image and then simulated it into a three-dimensional face model. Then, based on this face model, a person to be tested is identified. Face image, and the reason for the high false positive rate of this technique is 3 M364920. The complexity of the ambient light used in practice, especially under non-infrared lighting conditions, means the number of light sources, the degree of light source, the color of the light source, etc. Conditions, different from the original establishment of the surplus, the state of the photo, the environment, and the difference in the angle of illumination or the number, the face must be changed, meaning the position of the shadow, the position of the shadow and the file The data is the same, and the conventional technology uses a single camera to take a single image of the image. 'When using this light and shadow to simulate a three-dimensional image, and then compare it with the original image in the database, it is bound to be identified. People are not my own words. When someone uses the information of the person to know it, there is a possibility that the month b can be simulated as a The three-dimensional image of the file is similar, causing the imposter to be misjudged as a 'this is a serious problem in security'. It can be seen that the large-face recognition system caused by the conventional technology encounters the problem of declining recognition performance, which is ten. The existing face recognition system is really applicable. Most of the research work is to improve the existing visible system to reduce the impact of environmental lighting to make some progress, but the results are not good. Another technical problem is that people are very difficult because of the complexities of the face. These factors make the face a very challenging topic. The above-mentioned conventional techniques are mostly based on two-dimensional photos or dynamic video sequencing ί image processing technology; however, two-dimensional other changes are not able to solve the above-mentioned question itself. The reason is that the face of the face is π ί 2 Γ, and the two-dimensional photo is the heart of the three-dimensional surface, and the result of the projection is that the lack of the image of the ideal ring illumination angle is not the same as the light and shadow. Two-dimensional image two-dimensional image; When, the result of the shadow is the most dangerous part of the most dangerous human tires 4 M364920 will lose some of the method of the largest difference in the description of the structure of the face, etc.. Through the process of misjudgment The effect of illumination-independent and basic characteristics of the face, so that it can be seen that 3D people are currently in this field of information, in this process - dimensional recognition and traditional information can be better described and between points The better solution to the identification dimension face model has the ability to correctly reflect the two-dimensional structure without the representation of the 'face features. It can be a good solution to the bottleneck. Important information. Use three in the The comprehensive information of the depth information of the face features can be problematic, and at the same time, due to the characteristics of the three-state irrelevance, the face is mainly formed by the relatively stable face model recognition method. In order to solve the different changes in face pose and expression, it is mentioned that the ^3 laser scanner is used to capture the three-dimensional coordinates of the molded object and f = the dimensional model; or it is a structured light camera to fix the object on the object. Style stripe, then calculate the position of the object and construct the 3D model according to the shape of the stripe; or use the face image sequence = 2D, model. Among them, there is a technique for constructing a 3D face = using a single or multiple faces The image is used to construct a three-dimensional face model of the molded person. The universal three-dimensional face model is prepared in advance, and the general three-dimensional image is adjusted according to the position of the face feature point interactively, semi-automatically or automatically captured from the two-dimensional image. The shape of the face model. According to the feature points of each organ in the face image, the important feature points of the 3D model are reconstructed, and the relationship between the triangular vertices of the general 3D face model is added. The texture and mapping technology of 2D face images, and the establishment of a personal 3D face model. However, the virtual 3D face model with a single face shadow lacks face depth information. In addition, the 'light source is one that affects face recognition. The important factor of the item. The description of the above ^ has already explained that with the change of light, the information of the same face may change greatly, often causing the difference of the feature, so that the identification of the person is not the person, but not the person. The cost of 5 M364920 people poses a great threat to safety. In addition, if the three-dimensional laser scanner is used to extract three-dimensional human η. ^ This device system is affected by the ambient light, but the calculation is complicated. The degree is very high, and the laser light does not make the use of the actual system to meet the needs of the actual system...., the method of the external line or the far-infrared image is susceptible to the influence of the ring and the state of health, so that the change is obtained. 'In the actual application system, the sex month b is + good. r face a large face recognition in the face of the face has a literary single 茫 two silly times ^ two-dimensional face model of the practice of the face system is mostly used for multi-angle identification ', Virtually rely on a positive image to do the simulation, lack of depth information, ^ when the three-dimensional face is simulated, when the angle of view is changed: face angle)), it can be found that the face image will be long. - The most common kind of emotions: Skewed, or the shape of the nose is highly flawed' and one of the important five senses is misunderstood, -曰!J i's mistake, and the reason for the misjudgment here is because of the lack of people Face depth information. Y 疋 马 lack
$於上述人臉識別存在二大問題，本專利 ifίί!'搭配三維(3D)人臉識別裝置的方式S =實際拍攝的三維人臉影像進===模 $中=人=形狀模型則是以預先定義的特徵2樣 經過統計後得到。在某種局部點模型匹2 基礎上，利肖統計模型對待識別❸人臉的形進己的 6 M364920 斂到每p沾轉化為一個優化的問題，並期望最終收 ί ϊ ί人臉形狀。可以概分為兩大部分··註冊 ί tlf人ef階段。註冊階段提供人臉彩色數位照片 形狀模型，並將兩者結合起來而成人 番成山虽使用者走到紅外線光源之3D人臉識別裝 置，，由立體視覺取像平台進行拍攝，直接取得使 i ί Ϊ數位照片與三維人臉形狀模型，並進而將兩 者、〜a而成為三維人臉面具，意即可將人臉特徵加 入到3D人臉形狀模型上。識別階段工作在紅外線光 源照^下，不需要特殊照明條件，即可進行3D人臉 形狀模型識別。3D人臉識別系統可分為四個部份： (1)，入欲識別的人臉後，以適合的3D人臉形狀模 型尋找臉部區域。（2)代入事先以統計建立的3D人 臉形狀模型中。（3)在新的角度合成出所估測的人 臉。（4)與建立註冊資料庫中的3D人臉形狀模型 行識別。 ' 【新型内容】 本案的創作目的即在於發展出一種不受到環境 光源變化的影響且具有更高準確度的人臉識別f 置’因此’為了達到上述之目的，本創作提供一種 紅外線光源三維（3D)人臉識別裝置，其特徵^於包 括一立體攝影機’更包括一左攝影機與一右攝影 ，機’而該左攝影機用以取得一左影像，該右攝影機 用以取得一右影像；一紅外線光源，對所述立體攝 、影機提供照明；以及一影像處理器，與該立體攝影 機電連接’匹配該左影像與該右影像而形成一三^ 影像。 一 如上所述的紅外光源 紅外線發光二極體所構成 如上所述的紅外光源 其中該紅外線光源是由 其中該紅外線發光二極 7 M364920 體係組成陣列。 如上所述的立體攝影機’其中該立體攝影機的 感光元件係選自電荷耦合元件（CCD)或互補式金屬 氧化物半導體（CMOS)。 如上所述的裝置’其中該立體攝影機係透過乙 太網路線將所攝得的影像傳送到該影像處理器。 a如上所述的裝置’其中該影像處理器是設置於 一監控中心之内，而該影像處理器更與—顯示器連 線。 如上所述的裝置’其中該紅外線光源係設置於 該立體攝影機旁’並朝向該立體攝影機的視角方向 投射光線。 如上所述的裝置’其中該紅外線光源係圍繞該 立體攝影機而設置。 、為了達到上述之目的，本案還提供一種紅外線 光源三維人臉識別方法，包含以下步驟：使用預先 定義的特徵點及樣本影像經過統計後得到的人臉形 狀模型，與實際拍攝的三維人臉利用立體視覺匹配 進行三維重建’進行三維人臉識別。 如m所述’其中人臉形狀模型包含至少複數個 人臉特徵點所構成。 如前所述’其中人臉形狀模型為平均人臉加上 人臉外形模型產生的變異量；當人臉外形模型產生 的變異量為零時’人臉影像等於平均人臉。 如前所述’其中人臉形狀模型為具有人臉的平 移量、旋轉角度及縮放比例特性。 如前所述的立體視覺’其中包含至少兩部攝影 機構成的立體視覺架。為左右攝影機的投影中心的 連線的距離為基線距。空間中某點在左圖像和右圖 像中產生相對應的座標。計算某點在左右兩個攝影 8 M364920 機像面上的相應點，並且透過攝影機校正獲得攝影 •機的内外部參數’就可以確定這個點的三維座標c/ 如前所述’其中3D人臉識別包含具有輸入欲識 -別的人臉後’以適合的3D人臉形狀模型尋找臉部^ 域。 如前所述’其中3D人臉識別包含具有代入事先 以統計建立的3 D人臉形狀模型中。 如前所述，其中3D人臉識別包含具有在新的角 .度合成出所估測的人臉。 如前所述’其中3D人臉識別包含具有與建立註 修冊資料庫中的3D人臉形狀模型進行識別。 【實施方式】 請參閱圖2 ’為本案紅外線光源3D人臉識別裝 置的應用示意圖。其中’紅外線光源3D人臉識別裂 置在此應用上通常亦可稱為一個系統，意即若以^ 統觀之，本創作亦是三維人臉識別系統。本創作包 含立體攝影機Π、紅外線光源丨2和傳輸設備丨6(可 為乙太網路）等元件所構成。當人臉1〇以面向立體 影機1 3，利用安裝在立體攝影機丨3周圍的紅外線光 源1 2提供正面方向照明，將人臉i 〇及其上的特徵 例如眼部11的影像經立體攝影機13擷取並輸入到 監控中心1 7。紅外線光源丨2在任意環境光照下都是 清晰的，y建構不受環境光影響且高度準確的三 人臉，^系統，提供了良好的影像資料。三維人臉 ί ϋ f ί 5進一步對影像進行處理，以消除距離 逖近和頭邛安態所帶來的雜訊影響。而得到三維人 臉，ί的ί法在於本創作使用了立體攝影機13，其 更= ~攝影機131與—第二攝影機132，通常 攝影機呈左右排列，一如同樣呈左右排歹二艾 9 M364920 胜® ==圖3 ’為本創作的紅外光源三維人臉識別 形狀模型重建與人臉識別過程。其中第 夕，衫ΐ 1與第二攝影機132除了設置位置不同 卜’衫像的拍攝及其後的處理是相同的，首先是 ί人形狀模型模式，之後是人臉特徵定 ί11Γ疋以紅外線作為光源，因此人臉上的反差 ^為，，的就屬人眼’故本創作通常即以人 臉上定位的特徵，當人眼被定位後才ί =對人臉其他的特徵的定位，之後則是人臉角度計 iΐ述人臉角度計算是指人臉上各個特徵點之間 、计鼻，，常是相鄰的兩個特徵點進行計算。當人 臉角度計算完成後，即把此結果與一個預先定^ ^臉形f模型進行特徵擬合。於此之前所述的各個 〔，均是對第一攝影機131與第二攝影機132各自 拍攝到的影像所進行的步驟，而在此之後的立 像，配步驟，則是對於此二攝影機各自擬合的結^ ^行立體影像之匹配。之後，由於各攝影機是動態 ，影，因此可以得到一連串的立體影像，並且再^ 這些立體影像進行三維特徵點的標示，以利後來 步驟，即哥找影像的接合位置，此處所述的影 合位置係指各個立體影像之間用以相互接合的位 置，當各個立體影像接合成功之後，即建立了三 人臉模型’並將之儲存於資料庫内。 、 二維人臉識別方法，係使用預先定義的特徵點 及樣本影像經過統計後得到的人臉形狀模型，與每 際拍攝的三維人臉相比較，以進行識別的方法了 ^ 某種局部點模型匹配的基礎上，利用統計模型對 識別的人臉的形狀進行約束，從而轉化為一個 的問題’並期望最終收斂到實際的人臉形狀。可以 概分為兩大部分：註冊及識別階段。註冊階段提供 10 M364920 ί ί ί色ί位照片和3D人臉形狀模型，並將兩者結 ^ A而^ 3D人臉面具。當使用者走到紅外線光源 臉識別裝置時，由立體視覺取像平台進行拍 直^取得使用者的數位照片與三維人臉形狀模 ί收Τ ί而將兩者結合而成為三維人臉面具，意即 二—Ρ二2徵加入到3D人臉形狀模型上。通常，言主 ，Ϊ t Ϊ個作業程序、手段，可由監控中心透過 ί i ίί行控制。識別階段工作在紅外線光源照 特殊照明條件，即可進行3D人臉形狀 _ L 。3D人臉識別階段可分為四個步驟：（1)There are two major problems with the above face recognition, this patent is ίί! 'The way of matching three-dimensional (3D) face recognition device S = the actual captured 3D face image into === 模$中=人=shape model is It is obtained after statistics by pre-defined features. Based on a certain local point model 2, the Levi's statistical model treats the face of the face into the 6 M364920, which is transformed into an optimized problem and expects to eventually end up with a face shape. Can be divided into two major parts · registration ί tlf human ef stage. During the registration phase, a face color digital photo shape model is provided, and the two are combined to form a 3D face recognition device that the user walks to the infrared light source. The stereo vision image capturing platform is used to directly capture the i ί. The digital photo and the three-dimensional face shape model, and then the two, ~ a into a three-dimensional face mask, can be added to the 3D face shape model. The recognition stage works under the infrared light source, and the 3D face shape model recognition can be performed without special lighting conditions. The 3D face recognition system can be divided into four parts: (1) After entering the face to be recognized, look for the face area with a suitable 3D face shape model. (2) Substituting into the 3D face shape model established in advance by statistics. (3) Synthesize the estimated face from a new angle. (4) Establish a 3D face shape model in the registration database. '[New content] The purpose of this case is to develop a face recognition that is not affected by changes in ambient light sources and has higher accuracy. Therefore, in order to achieve the above objectives, the present invention provides an infrared light source three-dimensional ( 3D) a face recognition device, characterized in that it includes a stereo camera 'including a left camera and a right camera, and the left camera is used to obtain a left image, and the right camera is used to obtain a right image; An infrared light source for providing illumination to the stereo camera and the camera; and an image processor electrically matching the left image and the right image to form a three-image. An infrared light source as described above is composed of an infrared light emitting diode as described above. The infrared light source is an array of the infrared light emitting diode 7 M364920 system. The stereo camera as described above wherein the photosensitive element of the stereo camera is selected from a charge coupled device (CCD) or a complementary metal oxide semiconductor (CMOS). The apparatus as described above wherein the stereo camera transmits the captured image to the image processor via an Ethernet route. a device as described above, wherein the image processor is disposed within a monitoring center, and the image processor is further connected to the display. The apparatus as described above, wherein the infrared light source is disposed beside the stereo camera, projects light toward a viewing angle of the stereo camera. The apparatus as described above wherein the infrared light source is disposed around the stereo camera. In order to achieve the above purpose, the present invention also provides a three-dimensional face recognition method for infrared light source, which comprises the following steps: using a pre-defined feature point and a face shape model obtained by statistical sampling of the sample image, and utilizing the actually captured three-dimensional face. Stereoscopic matching for 3D reconstruction's 3D face recognition. As described in m, wherein the face shape model includes at least a plurality of face feature points. As described above, the face shape model is the variation of the average face plus the face shape model; when the face shape model produces a variation of zero, the face image is equal to the average face. As described above, the face shape model has the characteristics of the amount of translation, the angle of rotation, and the scaling of the face. The stereoscopic vision as described above includes a stereoscopic vision frame composed of at least two cameras. The distance to the line connecting the center of the left and right cameras is the baseline distance. A point in space produces a corresponding coordinate in the left and right images. Calculate the corresponding point of a point on the left and right sides of the 8 M364920 machine image, and obtain the internal and external parameters of the camera through the camera correction ' to determine the three-dimensional coordinates of this point c / as described above' where 3D face After the recognition contains the input ideology - after the face is selected, the face is searched for by the appropriate 3D face shape model. As described above, '3D face recognition includes a 3D face shape model that has been previously created by statistics. As previously mentioned, where 3D face recognition contains a face that is estimated at a new angle. As previously described, '3D face recognition includes recognition with a 3D face shape model in the build library database. [Embodiment] Please refer to Fig. 2' for the application of the infrared light source 3D face recognition device of the present invention. The 'infrared light source 3D face recognition cracking is also commonly referred to as a system in this application, which means that if the subject is viewed, the creation is also a three-dimensional face recognition system. This creation consists of components such as a stereo camera, an infrared light source 丨2, and a transmission device 丨6 (which can be an Ethernet). When the face is facing the stereoscopic camera 13 and the front side illumination is provided by the infrared light source 12 mounted around the stereo camera 丨3, the face i 〇 and the features thereon such as the image of the eye 11 are passed through the stereo camera. 13 Capture and input to the monitoring center 1 7. The infrared light source 丨2 is clear under any ambient light, and the y is constructed with a highly accurate three-faced, ^ system that provides good image data. 3D Faces ί ϋ f ί 5 Further processing the image to eliminate the effects of noise caused by close proximity and head 邛 state. In order to obtain a three-dimensional human face, the ί method is that the creation uses a stereo camera 13, which is more like a camera 131 and a second camera 132. Usually, the camera is arranged side by side, as if it is also left and right, and the second Ai 9 M364920 wins. ® == Figure 3 'This is the creation of an infrared source 3D face recognition shape model reconstruction and face recognition process. In the first day of the eve, the shirt 1 and the second camera 132 are the same except for the position of the shirt. The first shot is the same as the first shot. The first is the 形状 human shape model mode, followed by the face feature ί11Γ疋The light source, therefore, the contrast on the human face ^, is the human eye. 'The original creation is usually the feature of positioning on the human face. When the human eye is positioned, ί = the other features of the face are positioned. It is the face angle meter i. The calculation of the face angle refers to the calculation of the two feature points between the various feature points on the human face, counting the nose, and often adjacent. When the face angle calculation is completed, the result is fitted to a pre-determined ^^ face f model. Each of the previously described steps is a step performed on the images captured by the first camera 131 and the second camera 132, and the subsequent image, with the steps, is fitted to each of the two cameras. The knot ^ ^ line stereo image matching. After that, since each camera is dynamic and shadow, a series of stereoscopic images can be obtained, and these three-dimensional images are used to mark the three-dimensional feature points, so as to facilitate the later steps, that is, the joint position of the image, the image described here. The position is the position between the three stereo images for mutual engagement. After each stereo image is successfully joined, the three-person face model is established and stored in the database. The two-dimensional face recognition method uses a pre-defined feature point and a face shape model obtained after the sample image is statistically compared with the three-dimensional face photographed each time to identify the method. Based on the model matching, the statistical model is used to constrain the shape of the recognized face, which translates into a problem 'and expects to finally converge to the actual face shape. It can be divided into two major parts: the registration and identification phase. The registration stage provides 10 M364920 ί ί color ί photo and 3D face shape model, and the two are combined ^ A and ^ 3D face mask. When the user walks to the infrared light source face recognition device, the stereoscopic image capturing platform performs a straight shot to obtain the user's digital photo and the three-dimensional face shape mode, and combines the two to form a three-dimensional face mask. This means that the second-Ρ2 2 sign is added to the 3D face shape model. Usually, the main program, the operation program and the means can be controlled by the monitoring center through ί i ίί. The recognition phase works in the infrared light source under special lighting conditions, and the 3D face shape _ L can be performed. The 3D face recognition stage can be divided into four steps: (1)
j 2 =二，的人臉後，以適合的三維人臉形狀模型 寻找臉4區域；f 9彳你X 由.k巧（）代一通用三維人臉形狀模型 诸二a m 的角度合成出所估測的人臉；以及（4)與 Ψ i*Μ ^料^中的人臉形狀模型進行識別。其 餘斗二古々用二維人臉形狀模型，係本創作事前以 統计的方式所建立者。 # 士 2 ΐ，ί既存於市面上的各種人臉識別系統的 維人瞼二刑就在於i本創作的資料庫中記錄的是三 :ϋί 」而不是儲存多姿態的人臉二維照片。 ΐ ; Ϊ ΐ Ϊ机的投影變化可以輕易的解決人臉識別 Ιίίίϊ問題，利用物體幾何原㉟，排除大部分 別的影響’而且由於採用的紅外線影 ΐίΐ:ΐ明暗程度無關，可以解決人臉識別 “識；ί:環境光照強度的影響，達到較理想的 似$ = 1 ^人臉辨識研究大多以二維影像之間的相 的s #交ΐ。太ϊ=ΐ i度主要是建立在灰階強度值 飧盔士认；！本專利可犬破這個限制’因為以二維影 識：此係因ΪΪΓίΪΐί不同人臉角度的影像辨 局田規，丁、的角度不同之時，同一個人臉 11 M364920 卻會有不同的灰階資訊，故為了解決這個問題，需 要將三維人臉模型的資訊引入人臉影像辨識中。包 含有一個人臉偵測演算法、一個統計的人臉形狀模 型、一個建立三維的人臉統計模型方法及一個可靠 的人臉影像比對法。最後’形成一個紅外光源三維 =別裝置之人臉形狀模型重建與人臉識別過程如之 前的圖3所示者，為此，本案亦發展出以下幾個在 人臉辨識技術中的主要特色： ^ 1一個供系統管理人員在監控中心使用的管理 平台，是一應用軟體，主要進行三維人臉資料特徵 ，的擷取和儲存等功能，並用以操作本案的紅外線 土源三維人臉識別裝置，由於通過立體視覺攝影機 擷取原始的人臉三維模型而得到的三維資料座標可 ΐ不，為了便於計算使用，要進行座標的正規 =，對原始二維模型的平滑化及座標轉換， 人臉的三維特徵點的選取工作亦可：j 2 = two, after the face, look for the face 4 area with a suitable three-dimensional face shape model; f 9 彳 you X is synthesized by the angle of .k () and a general three-dimensional face shape model The face is measured; and (4) is identified with the face shape model in Ψ i*Μ^^. The two-dimensional face shape model of Yudou Ergu was established in a statistical manner beforehand. #士 2 ΐ, ί The various types of face recognition systems that exist in the market are based on the fact that the document is recorded in the database of i: ϋί ” instead of storing two-dimensional photos of faces with multiple poses. ΐ ; Ϊ 投影 The projection change of the machine can easily solve the face recognition Ιίίίϊ problem, using the object geometry 35, to eliminate most of the other effects' and because of the infrared image ΐ ΐ ΐ ΐ ΐ ΐ 无关 无关 人 人 人 人 人 人 人 人 人 人 人 人"Knowledge; ί: the effect of ambient light intensity, to achieve a better than $ = 1 ^ face recognition research mostly with the phase s # intersection between two-dimensional images. Too = ΐ i degree is mainly based on gray The intensity value of the order is recognized by the helmet; this patent can break the limit of the dog's because of the two-dimensional image recognition: this is because the image of the face of the face is different, 11 M364920 has different gray level information, so in order to solve this problem, the information of 3D face model needs to be introduced into face image recognition. It includes a face detection algorithm, a statistical face shape model, and a Establish a three-dimensional face statistical model method and a reliable face image comparison method. Finally, 'form an infrared light source three-dimensional = other device face shape model reconstruction and face recognition The process is as shown in Figure 3 above. To this end, the following features are also developed in the face recognition technology: ^ 1 A management platform used by system administrators in the monitoring center is an application software. It mainly performs three-dimensional face data feature capture, storage and other functions, and is used to operate the infrared soil source three-dimensional face recognition device of the present case. The three-dimensional data coordinates obtained by capturing the original human face three-dimensional model through the stereoscopic camera can be ΐ No, in order to facilitate the calculation and use, the coordinates of the coordinates should be normalized =, the smoothing of the original two-dimensional model and the coordinate transformation, the selection of the three-dimensional feature points of the face can also be:
對特徵點進行標示。令邱轳千社击从 e 十D 資料庫di = : ^後’記錄入電腦 者的增加、刪除、修改等常用資料庫管理功能。 (2)一個人臉自動化識別方法。首先，透禍紅冰 ίϋ配合立體攝影機對人臉影像進行分析，自2 刭，=人臉存在位置並從影像中將其分割出來。 俨_臉^，使用人臉形狀模型的特徵點 = Γ H 右攝影機分別可得到立體影像特ϊ ί 根據主要的特徵點位置，計算出、‘” 12 1 ί 21! 丁讀取，讀取後建構出人臉的三錐 的平面投聲 3f的士，模型。當三維模型 通過多、中人臉的旋轉角度相同時， 過夕組特斂向置、分類器判別及臨 M364920 法實現人臉識別。 (3)使用資料庫之資嵙絲+ ，石 能力’實現系統需要使用到J性和大量資料、： ⑷不論是管理平台二'查自詢檢索二 ^ 資料庫的讀取都通過資料=^動識別系統，對人臉 護及操作。 、庫元成進行資訊讀取、維 若使用本專利提出安驻 外光源提供正面方向昭明裝在立體攝影機周圍的紅 任意環境光照下都是清aj的以發現紅外線光源在 響且高度準確的人臉景;^的，可得到不受環境光影 為了克服單張影像欠缺、听由〜 a 預先定義的特徵點及樣賢訊的問題，利用 臉形狀模型與實際拍攝的象給莖過統计/後侍到的人 法。在某種局部點模型匹配一、、，^臉進仃識別的方 型對待識別的人臉的形狀進以巧上2 m模 個優化的問題，並期望最:束，從而轉化為一 狀。並採用雙眼立體影像取架^收&气si際的人臉形 攝左右影•，之後並、經人臉=彡2影機分別拍 三維（3D)人臉模型。 ^狀极型處理可以得到 將人臉形狀模型定義為— 整模型參數使能量函數最，個=里函數，透過調 ^ ^ I. a « „ ^ f ^ 49 制，從而將形狀的改變限制 / ^的調即加以限 首先，對-組標有特徵變點限圖隹 進行灰度建模，然後在搜索像的形狀和局部 姿態參數，從而使形狀達到化斷调節形狀和 請參閱圖4(a)、圖4(b) 為本創作人臉形狀於應用時;」）、與圖4(d)， 練人臉影像的形狀變化，對不同狀為了研究訓 M364920 進行比較’應先對這些人臉影像進行校正，苴 形狀進行旋轉、縮放和平移使其盡可能的蛊臭^ 狀接近。然後，對校正後的形狀資料進行 >主$量‘ 析（Principal Component Analysis，PCA)，過程如下. (1)計算校正後的形狀向量的共變異矩陣： . Σ, ；^Σ(ί)(ί)7 (2) 透過下式計算共變異的特徵值八认，A，..· (3) 取相應的特徵向量並正規化，記户=( 與較大的特徵值Α對庫的特 &，仏，…^) 0 形狀模型的變化，這二量，代表較重要的 個特徵向量表示為： 白了以用刖 甘 i x = x + Pb 户疋已訓練好的模划，/) a , 個新臉部影像x等於疋-個向量。- ΐίfi的變異量」。當△為零時，人臉ί ) 平均人臉无。當炎參士 才人臉影像X等於 種表情或姿離。6 __ Α’、ν % ’人臉影像可能正在做某 數。以下^圖Λ±·.产y是控制前t個模式的ί 中’圖4(a)為基本的人臉f 7步的說明’其 從左到右相應AH形狀，如圖4_示， ^ = (-3^0,0,---,0/ > ύ~(〇3 ίΓη Γ、0==(3V^，〇,〇,...，〇)r、 圖x人所有0 =户(卜无）。 人臉擬合過程'中f ^莫型之控制參數示意圖，在 變外部參數中&匕迭：迴圈的過程，不％; 至形狀變不大， ^轉角度及縮放比例，直 了 ^為此時的形狀即所要搜索的| 14 M364920 像形狀 的人臉 b ： 數。 P : 。1過人臉形狀模型產生人臉「擬合 最〜々擬合有幾項控制參數可供調口塾」广入 式〇)，影響人臉影像本體X，屬於内 置，屬於中心點放置在輸入影像的座標位 外部參數動’“象放置在輸入影像之尺寸比例’屬於 差条L主„像中心線相對於輸入影像座標軸之偏 差角度，屬於外部參數。 神又偏 噴敕可，"I列式（4)表示人臉影像χ經過各項參數的 調整，逼近輸入影像致的 f^P^e)^\Y~X\, 〇ptimize(Y,X(bXp^e) = argmm(f) (4) 赵 ΐ /是人臉影像與輸入影像的相似誤差函 馬僮ί ϊ化擬合如式（4)，不斷的調整參數直到人臉 與輸入影像達到誤差最小化為止，此時& = 即是需要的人臉特徵資訊。 n里 言月亡^圖5(a)、圖5(b}、圖5(〇為立體攝影機 響 ^置示意圖。其中，立體攝影機實際上可說是為 =實踐立體視覺模型而設置的硬體，其中圖5(a)以 二維的視角說明第一攝影機131與第二攝影機132 置的方式與世界座標的關係；而圖5(b)則顯示兩 4攝影機光軸互相平行，向量/和向量/，垂直於基準 線办’攝影機視軸角度（pan)，= 0。請參閱圖5(c)，第 二攝影機131繞著7軸旋轉、而第二攝影機132繞 著广軸旋轉’它們的光軸交在同一點，此點稱為凝 才見點（Fixation p〇int)，產生視軸角度夕。第一攝影機 先袖和向量/形成一個視軸角度^為/的右 邊。同理’第二攝影機光軸和向量/，形成一個視軸 15 M364920 角度夕2，’2>0為’的左邊。兩部攝旦〈 一 可以利用攝影機外部參數得至丨 ^ 相隔基準線6 侍刿下列方程式： (Γ_Τ) Μ: A = cos-%，，久 (5) 式中，i是單位向量，7^和r，分別 移矩陣，犀和 <分別是左右攝影 =f右攝影機的平 向量。 疋轉矩陣的第一列 請參閱圖6(a)、圖6(b)。其中阁< 體攝影機的立體場景座標示音圖、_ 6(a)為本創作立 的針孔模型示意圖。假設第^一搌:圖6(b)為本創作 影機1 32的俯仰（tilt)角相同，二Ί 13 1與第二攝 型可以用圖6(b)來表示，並中厂^攝影機的針孔模 為（Hz)，而分別在第一攝、影？ ^測點户的座標 132内部的感光元件上呈像， 一與第二攝影機 二為（Xl，少1)、第二呈像點P2的座ρ二f像點Pl的座 j 6(a)中的各個符號集中的略 ^為以下將 ，影機（左攝影機）的光學中心，〇兒】笛其中，〇為第一 ^ π u，）為第」影機)的攝影機座 度^為第二攝影機的：軸角巧： 儿體攝影機的基準線； J沉孕由冉度、為 21票：則是第-攝影機影像平面;點影？的影像平面Mark the feature points. Qiu Qianqian was attacked from the e 10 D database di = : ^ after 'recorded into the computer's common database management functions such as addition, deletion, modification. (2) A method for automatic recognition of faces. First of all, the stereo image is analyzed with a stereo camera. From 2 刭, = the face exists and is segmented from the image.俨_face^, using the feature point of the face shape model = Γ H The right camera can get the stereo image feature ί According to the main feature point position, calculate, '' 12 1 ί 21! Ding read, after reading Constructing a three-cone flat-projection 3f taxi, model of the face. When the three-dimensional model passes through the same angle of rotation of the multi- and medium-faced faces, the Essence group has a special convergence, the classifier discriminates and the M364920 method realizes the face. Identification (3) The use of the database of the resources of silk +, stone ability 'to achieve the system needs to use J and a large amount of information,: (4) whether it is the management platform two 'chasing self-inquiry search two ^ database read through the data =^Motion recognition system, for face protection and operation., Ku Yuancheng for information reading, Wei Ruo use this patent to provide an external light source to provide a positive direction. Zhao Ming installed in the red around the stereo camera, any environment lighting is clear Aj's to find the infrared light source in the loud and highly accurate face scene; ^, can get the problem of being free from ambient light and shadow in order to overcome the lack of single image, listen to ~ a pre-defined feature points and sample good, use the face Shape mode The method of statistic/post-serving with the actual image taken. In a certain local point model, the shape of the face to be recognized by the face, the face, and the shape of the face to be recognized is 2 m. Modeling an optimization problem, and expecting the most: beam, which is transformed into a shape, and adopts a binocular stereo image to take the frame and collect the image of the face shape of the gas, and then, after the face = 彡 2 The camera separately takes a three-dimensional (3D) face model. The shape-like processing can be used to define the face shape model as - the entire model parameter makes the energy function the most, one = the inner function, through the tuning ^ ^ I. a « „ ^ f ^ 49 system, so that the shape change limit / ^ tone is limited to the first, the - group marked with a variable point limit map 隹 gray modeling, and then search for the image shape and local pose parameters, thereby making The shape is changed to the shape of the adjustment and see Fig. 4 (a), Fig. 4 (b) is the shape of the artificial face in the application; "), and Figure 4 (d), the shape of the face image changes, Different shapes are compared for the research training M364920 'The face images should be corrected first, the shape is rotated, Zoom and pan to make it as close as possible to the odor. Then, the corrected shape data is subjected to >Principal Component Analysis (PCA). The process is as follows. (1) Calculate the covariation matrix of the corrected shape vector: . Σ, ;^Σ(ί) (ί)7 (2) Calculate the eigenvalues of the covariation by the following formula, A,..· (3) Take the corresponding eigenvectors and normalize them, and record the households = (with larger eigenvalues Α for the library) Special &,仏,...^) 0 The change of the shape model, these two quantities represent the more important eigenvectors expressed as: White to use the 刖 = = x + Pb 疋 疋 trained training, / a) A new facial image x is equal to 疋-vector. - 变异ίfi's variation." When △ is zero, the face ί ) has no average face. When Yan Shishi's face image X is equal to a kind of expression or posture. 6 __ Α’, ν % ’ The face image may be doing some number. The following figure Λ±·.Production y is the control of the first t modes ί 'Fig. 4 (a) is the basic face f 7 step description 'the corresponding AH shape from left to right, as shown in Figure 4 ^ = (-3^0,0,---,0/ > ύ~(〇3 ίΓη Γ, 0==(3V^,〇,〇,...,〇)r, Figure x people all 0 = household (Bu no). The face fitting process 'in the f ^ mo type control parameter diagram, in the variable external parameters & overlap: the process of the loop, not %; to the shape does not change, ^ turn angle And the scaling ratio, straight ^ is the shape of this time that is to be searched | 14 M364920 The shape of the face b: number. P : 1. The face shape model produces a face "Fitting the most ~ 々 fit a few The item control parameter can be used to adjust the 塾 塾 广 广 广 广 , , , , , , , , , , , , , , 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响 影响The deviation L main „ like the deviation angle of the center line relative to the input image coordinate axis, is an external parameter. God is sneezing, "I column (4) means that the face image is adjusted by various parameters, approximating the loss Image-induced f^P^e)^\Y~X\, 〇ptimize(Y,X(bXp^e) = argmm(f) (4) Zhao Wei / is the similar error of the face image and the input image Tong ϊ 拟合 拟合 如 如 如 如 如 如 如 如 如 如 如 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合 拟合(a), Fig. 5(b}, Fig. 5 (〇 is a schematic diagram of the stereo camera sound. Among them, the stereo camera can be said to be a hardware set for the practice stereoscopic visual model, wherein Fig. 5(a) The two-dimensional view illustrates the relationship between the first camera 131 and the second camera 132 and the world coordinates; and FIG. 5(b) shows that the optical axes of the two cameras are parallel to each other, vector/sum vector/, perpendicular to the baseline. 'Camera bore angle (pan), = 0. Referring to Fig. 5(c), the second camera 131 rotates about the 7 axis, and the second camera 132 rotates about the wide axis 'their optical axes intersect at the same point, This point is called the Fixation p〇int, which produces the visual axis angle. The first camera first sleeve and the vector / form a right axis angle ^ is the right side of the same. The second camera optical axis and vector /, form a visual axis 15 M364920 angle eve 2, '2> 0 is the left side of the two. Two parts of the film can be used to take advantage of the external parameters of the camera 丨 ^ separated from the baseline 6 Equation: (Γ_Τ) Μ: A = cos-%,, long (5) where i is the unit vector, 7^ and r, respectively shift matrix, rhinoceros and < respectively, left and right photography = f right camera flat vector . The first column of the twist matrix is shown in Figure 6(a) and Figure 6(b). The stereoscopic scene of the cabinet & body camera marks the sound map, and _ 6(a) is a schematic diagram of the pinhole model of the original. Assume the first one: Fig. 6(b) shows that the tilt angle of the creation camera 1 32 is the same, and the second 13 1 and the second image can be represented by Fig. 6(b), and the factory camera The pinhole mode is (Hz), but in the first shot, shadow? ^The image of the photosensitive element inside the coordinate 132 of the measuring household is imaged, and the second camera 2 is (Xl, 1 less), and the second image forming point P2 is the seat ρ2f image point P1 of the seat j 6(a) In the respective symbol sets in the following, the optical center of the camera (left camera), the 】 】 笛 其中 其中 其中 其中 其中 ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ ^ 第 第 第The second camera: the angle of the shaft: the baseline of the child camera; J is pregnant by the degree, 21 votes: is the camera-image plane; point shadow? Image plane
J在各攝影機平面(又 位置，Q點則是P '就是QP:乂；；離攝影機平面(X，Z)的高度：2 J Q』與兩讀影機的基準線办的距離，因此，1 16 M364920 為Q點與〇1货 ^ ❿為Q點鱼 攝影機影像平面原點之間的夾角、而 為3D場景、第二攝影機影像平面原點之間的夾角，p 立體攝影機μ固定點。請繼續參閱圖6(a)，為本創作 中立體場景座標示意圖。第-攝影機的鏡^ 攝影機和/又二第二攝影機的鏡頭中心為（M，0)。第一 θ 向量夾角為夕r第二攝影機和Γ向量夾角為 义，兩部攝影機視軸角度Α。由（） 為 算出三維場景點^神月度Α甶（Χ，凡（Χ，少）和Ζ可以計J is in the plane of each camera (again, Q is P' is QP: 乂;; height from camera plane (X, Z): 2 JQ" and the distance between the reference lines of the two readers, therefore, 1 16 M364920 is the angle between Q point and 〇1 goods ^ ❿ is the angle between the origin of the Q point fish camera image plane, and the angle between the 3D scene and the origin of the second camera image plane, p stereo camera μ fixed point. Continue to refer to Figure 6(a), which is a schematic diagram of the coordinates of the stereoscopic scene in the creation. The lens center of the camera-camera and/or the second camera of the camera is (M, 0). The angle of the first θ vector is 夕r. The angle between the two camera and the Γ vector is the same, and the angle of the camera's boresight is Α. From () to calculate the three-dimensional scene point ^ 神月度Α甶 (Χ,凡(Χ,少) and Ζ can count
—=tan(^ +S) + tan(/?2 + Φ) (6) 式中，<D = _tan-i((x’ —心/㈣，/和尸代 表以pixels表示的焦距。 在獲得Z後，户點的X和r位置可以計算如下： 尤=Z 1；311(成 +5)或尤=办—Z tan(y02 + φ) (7' 和 y 一 (y 1n)ZCOS⑹ _ (〆-X)|Z|C0S(O) |/v|cos(^+^) ΚΙ cos(/?2 + Φ) (8) 就習知技術而言，獲得一個人在不同情況下的 臉部影像主要有兩種方式：第一，拍攝並存槽一個 人在各種情況下的臉部照片，弟二’拍攝一張標準 的臉部照片，經由電腦軟體處理的方式獲得多種光 線和裝扮的臉部影像。然而一如前述，此二種習知 技術均有缺點。而本案則疋透過人臉形狀模型及立 體視覺技術，將左右人臉影像取得的人臉形狀模型 經由立體匹配後可構成三維（3 D)人臉模型影像，此 三維人臉模型可以定義個人的各種不同臉部姿態， 具有人臉深度貢说。 17 為輔 識系 人臉 用者 方便 ，經 貌， 變4匕 源搭 照變 義的 模型 局部 的人 題， 源能 人臉 深度 ❿ M364920 肋人在於推測人臉正面的影像，以作 统，厂3美二ΐ識的工具。目前的自動人臉辨 i^ ϊ ϊ 的人臉影像，對於非正面的 則效果較差或無法使用。如此的設計限制了 必須正對攝影機入鏡，大大的限制了使用者的 =$系統的適用性。本案經由事前的統計學習 由二維人臉形狀模型資訊來預測重建正面的 以方便後續分析辨識的處理。可選擇減輕表 所產^的影響，有助於人臉辨識時的準確性。 一細上所述’本案創新之處在於以红外線 配三維人臉識別裝置的方式，能解決環境光 化、人臉姿態及表情的不同變化。使用預先 特徵點及樣本影像經過統計後得到的人臉形 與貫際拍攝的3D人臉進行識別的方法。在某 點模型匹配的基礎上，利用統計模型對待^ 臉的形狀進行約束，從而轉化為一個優化的問 亚期^最終收斂到實際的人臉形狀。红外線光 解決環境光照變化干擾、人臉形狀模型能提高 搜尋精準性，而立體視覺能重建三維人臉模型 資訊’屬於嶄新的技術。 此外，本案較習用技術更為優秀之處在於， 著光線的變化，同一張人臉的資訊可能產生很大 變化，往往造成人臉特徵的差別，當光線不足 j用可見光攝影機的習用技術無法解決這個問題。 若使用本案所提出的紅外線光源則可忽略環境光照 變化干擾，由紅外光源對人臉拍攝得到的影像可4以、 看出人臉眼睛部份會發亮，可以鮮明的在立體攝$ 機上呈像，並進而增強人臉識別率。加上立體视= 架構可以在一次的拍攝作業中，就建立起人二 模型並獲得景深資訊，比起以單張人臉影像所^擬 18 M364920 出的二維人臉模型的習一 -臉深度資訊普遍不準〗，用技術的人 的清晰拍攝效果、A A f觀本案透過紅外線光源 •度資訊，可以更有& Μ #體攝影機之準確的人臉深 而在產業3=;人”工作。 產業上，就辨識區域方面，可運用於安全監控 i眼睛貞擷取臉部特徵1 及臉部其他區域：皮；：二卜3的,：和相對位置 未經允許人員進入，上亨f，糸統之功能’防止 的進出次數、、時間長》疋σ賞T系統用以紀錄員工 此外亦可輔助駕π現的地點場所等， ;方ϊ 以下參考附圖說明本創作實 【圖式簡單說明】 =1.，為傳統人臉識別裝置之設計架構圖（先前技 為本創作紅外線光源3D人臉識別裝置示意 圖2 圖； 圖3，為本創作的紅外光源三維人臉識別裝置之 人臉形狀模型重建與人臉識別過程； 圖4(a)、圖4(b)、圖4(c)、與圖4(d)，為本創作 人臉形狀於應用時的示意圖； 圖5(a)、圖5(b)、圖5(c)為立體攝影機的設置示 意圖； 圖6(a) ’為本創作立體攝影機的立體場景座標示音 圖；以及 圖6(b) ’為本創作的針孔模型示意圖。 M364920 【主要元件符號說明】 10 ： 人臉 11 ： 眼睛 12 ： 紅外光源 13 ： 立體攝影機 1 3 1 :第一攝影機 132 :第二攝影機—=tan(^ +S) + tan(/?2 + Φ) (6) where <D = _tan-i((x' - heart/(4), / and the corpse represents the focal length in pixels. After obtaining Z, the X and r positions of the household point can be calculated as follows: 尤=Z 1; 311 (+5) or 尤=办—Z tan(y02 + φ) (7' and y (y 1n) ZCOS(6) _ (〆-X)|Z|C0S(O) |/v|cos(^+^) ΚΙ cos(/?2 + Φ) (8) In terms of conventional techniques, obtain a face of a person in different situations There are two main ways of imagery: first, taking a picture of a person's face in various situations, and taking a picture of a standard face, and using a computer software to obtain a variety of light and face images. However, as mentioned above, these two conventional techniques have disadvantages. In this case, the face shape model obtained by the left and right face images can be three-dimensionally matched by the face shape model and the stereoscopic vision technology (3) D) Face model image, this 3D face model can define various facial gestures of the individual, and has a face depth tribute. 17 For the auxiliary knowledge, the face is convenient for the user, the appearance, the change According to the model of the variant, the source of the face can be deep. 364 M364920 The rib is to speculate on the image of the front of the face, and to use it as a tool for the factory. The current automatic face recognition i^ ϊ ϊ The face image is not effective or unusable for non-positive. This design limits the need to face the camera, greatly limiting the applicability of the user's =$ system. The face shape model information is used to predict the reconstruction of the front to facilitate the subsequent analysis and identification. It is possible to reduce the impact of the table and to help the accuracy of face recognition. In the way of infrared three-dimensional face recognition device, it can solve different changes of ambient light, face posture and expression. The face shape and the 3D face of the continuous shooting are obtained by using the pre-feature points and the sample images. The method of identification. On the basis of model matching at a certain point, the statistical model is used to constrain the shape of the face, which is transformed into an optimized Q-phase. Converging to the actual shape of the face. Infrared light solves the disturbance of ambient light changes, the face shape model can improve the search accuracy, and stereoscopic vision can reconstruct the 3D face model information' is a new technology. In addition, this case is more practical than the conventional technology. The advantage is that with the change of light, the information of the same face may change greatly, which often causes the difference of facial features. When the light is insufficient, the problem can not be solved by the conventional technology of visible light camera. The infrared light source can ignore the interference of ambient light changes, and the image obtained by the infrared light source on the human face can be seen as a part of the human face, which can be brightly displayed on the stereo camera, and further Enhance face recognition rate. In addition, the stereoscopic view=architecture can establish a human model and obtain depth information in one shooting operation, compared to the one-face of the two-dimensional face model of 18 M364920 with a single face image. In-depth information is generally not allowed, the use of technical people's clear shooting effect, AA f view this case through the infrared light source • degree information, can be more & Μ # body camera accurate face deep in the industry 3 =; people" Industry. In terms of identification area, it can be used for security monitoring. Eyes capture facial features 1 and other areas of the face: skin;: 2 Bu 3,: and relative positions are not allowed to enter, Shangheng f, the function of the system 'prevents the number of times of entry and exit, long time' 疋 赏 赏 T system used to record employees can also assist in driving the location of the site, etc.; Brief description of the formula] =1., is the design of the traditional face recognition device architecture (previously the schematic of the infrared light source 3D face recognition device 2); Figure 3, the creation of the infrared light source three-dimensional face recognition device human face Figure reconstruction (Fig. 4(a), Fig. 4(b), Fig. 4(c), and Fig. 4(d) are schematic diagrams of the creation of the face shape of the creator; Figure 5 (a Fig. 5(b) and Fig. 5(c) are schematic diagrams of the arrangement of the stereo camera; Fig. 6(a) 'is the stereoscopic scene seat sound map of the creation stereo camera; and Fig. 6(b) 'is the creation Schematic diagram of pinhole model M364920 [Key component symbol description] 10 : Face 11 : Eye 12 : Infrared light source 13 : Stereo camera 1 3 1 : First camera 132 : Second camera
14 ： 顯示裝置 15 ： 訊號處理裝置 16 ： 訊號傳輸裝置 17 ： 監控中心 2014 : Display device 15 : Signal processing device 16 : Signal transmission device 17 : Monitoring center 20