經濟部中央標準局員工消費合作枉印製 4洲3 A7 B7 五、’發明説明（1 ) 本發明係有關於一種立體顯像系統及方法，特別係Printed by the staff of the Central Standards Bureau of the Ministry of Economic Affairs on consumer cooperation. 4 Continents 3 A7 B7 V. Description of the Invention (1) The present invention relates to a three-dimensional imaging system and method.
有關於一顏以影像處理法製作的追蹤式立體顯像系統及 方法V 在虛擬實境技術的發展過程_，為了使觀察者能夠 觀看到立體影像，一般均須要求觀察者配載特殊的觀看 裝置，例如眼鏡等，藉以分別傳輸不同的影像到觀察者 的雙眼。然而，配載這些觀看裝置卻會造成對觀察者的 束缚，同時觀察者的移動範圍亦受到限制。 為了改進此種缺點’在先前的技術中，另外也發展 出一種追蹤式立體顯像系統，藉著觀察者所在位置的資 訊，可減少利用時間多工與空間多工方式所必須做的聚 、焦重複》 應用在虛擬實境中的追蹤系統，目前有五種方式： (1)機械式’其係利用連桿裝置所組成，運用此種方式的 裝置會對使用者造成束缚，移動範圍也同樣地會受到限 制》(2)電磁式’利用三個接收天線，分別接收三組連續 發射的電磁波’利用九個電磁場強度的數據，計算使用 者頭部的位置與方向，但其延滯時間長，追蹤範圍較小， 不適用於使用者不斷移動的狀況，且容易受外界磁場的 影響β (3)超音波感應式，其原理類似於電磁式，即利用 三組超音波的傳輸時間差來作三角定位，但其準確度受 空氣密度變化，或因其他物體遮蔽，容易產生誤差。（4) 光學式，其利用雷射或紅外線作為光源，以特定的排列 方式安裝在周圍環境中，同時在頭載顯示器上，設置感 4 本紙G度適财賴家料（CNS Μ術M :1G X加公$ )—'— ----—~- 裝------訂 --‘---r (請先閲讀背面之注意事項再填寫本頁) 413793 經濟部中央標準局員工消費合作杜印裝 A 7 B7 五、發明説明（2 ) 測器，利用光電管產生的電流大小，和光線中心在感測 器上的位置，推算頭部的位置與方向，不過其追蹤範圍 受到限制， 在先前技術中，為了提供給使用者最自然的立體顯 像方式，發展出利用面孔探測的影像處理技術，決定觀 察者的面孔位置，藉以做為追蹤觀察者雙眼位置的依 據，引導立體顯像系統聚焦並投射影像至觀察者的雙眼 中。 有關於此種影像處理式的追蹤式立體顯像系統及方 法的先前技術，大致上可分為下列幾種方法： 一、神經網路式（neural network-based)，例如（l)S.H. Lin et al·發表在 Wewra/ Vol_ 8，pp. 114-132，Jan. 1997 的「Face Recognition/ Detection by Probabilistic Decision-Based Neural Network」，其採用 兩階段式的學習修正，先是locally unsupervised learning，每個子網路分別訓練，以EM法估算由影像中 之高頻成分組成的子空間參數，待此區域性的訓練完 成，便進入globally supervised training，再對由區域性 訓練中得到的邊界，依據數據作進一步的調整。（2)在S. Lawrence et al.的"Convolutional Neural-Network for Face Recognition " , Proc. IEEE Conf .on Computer Vision 仙£^尸<3"^«及6(^尽則'"〇«，卩卩.217-：222，1996_中，其結合自 組圖神經網路（self-organizing map network)及折積神經 網路（convolutional neural network)。自組圖神經網路可 (請先閲讀背面之注意事項再填寫本I ) 本紙張尺度適用中國國家標準（CNS ) A4規格（210X297公釐） 413793 a7 B7 經濟部中央橾隼局員工消費合作社印製 五、發明説明（3 ) 降低輸入維數（解析度92'112的灰階），使影像在拓樸空 間（topological space)上定量，而折積神經網路則對位移、 旋轉、大小變化以及變形等提供一個部份的不變性，在 該份報告中的辨識率較固有向量法佳°(3)在H. A. Rowley et al.的"NeuralNetwork-BasedFaceDetection’,，Proc· IEEE Conf. on Computer Vision and Pattern Recognition, 1996.中，其先以修正照明與柱狀圖等化作數據前處理， 再以各種視窗尺寸和各次樣本比例去偵測可能的面孔， 此方式以1050張訓練面孔影像訓練的濒神經網路，以130 張測試面孔，可達90.5°/。的辨識率。 二、以面容為基礎（Feature-based)的面孔辨識首先由 可變播的模板（template)裝配到五官上，萃取五官的幾何 形狀，再將之常態化（normalize)成標準形、·常態化的容貌 -- ' ^ 像再經由一組主要的思有向量分類。此一分類所用的 代表向量則可對比到五官，而不受到表情變化的複雜影 響。在 I.J. Cox et al·的"Feature-based Face Recognition Using Mixture-Distance", Proc. IEEE Conf. on Computer Paiierw1996 尹，以六個位能項目 包括各變數關係，在一個循環重複的模板裝配過程中， 逐步修正各參數，但是這個方法所得到的辨識率並不理 想。因此，在 M.A.Shackleton et al•的"Classification of Facial Features for Recognition", Proc. IEEE Conf. on 1991 中，採用 了混合距離技術（mixture-distance)，每張面孔由35個量 (請先閲讀背面之注意事項再填寫本頁) 裝· 111 本纸張尺度適用中國國家標準（CNS ) A4規格（210X297公釐） 經濟部中央標準局員工消费合作社印製 A 7 B7五、發明説明（4 ) 測的位置所推算的30個距離來代表，可達到95%的識別 率，不過其計算量太大。 三、以視覺為基礎的特徵空間，此方式假定所有可 能的面孔圖案均佔有一小的且變數確定的次空間。此方 式利用面孔影像的一個或多傰範例組中的資料群及主分 量逼近面孔圖案的次空間。\特徵空間的方式具有一個優 點，即其可處理封閉的狀況。然而此種方式僅被證實可 工作於非群集的背景中。相關的資料可參閱M.A. Turk et al.的1’Face Recognition Using Eigenfaces", Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1991 、 T. Darrell et al.的"Active Face Tracking and Pose Estimation in an Interactive Room", Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 1996 、及美國 專利第5,1 64,992號。此種方式使用特徵面孔（eigenfaces) 作面孔偵測，具有快速、簡單的特點，但受限於其偵測 需依賴訓練及測試影像間之各畫素的高度相關性。 在立體顯像系統中，需要追蹤系統前的正面主要觀 察者的面孔位置，在可接受的計算量之下，連續地輪出 觀察者雙目的位置，藉以作為立體顯像的標的空間。然 而在前述的先前技術中，面孔辨識與追蹤的系統存在有 計算量過重、必須龐大的人工處理及冗長的推導與循環 重複運算的收斂等問題。 有鑑於此，為了減少立體顯像系統中，時間多工與 空間多工所造成的重複聚焦，本發明之目的即在於提供 n Hu ^1.^1 ^^1· 1 I— I HI •士 —I u -- _ -1— I 、T - - - - (請先閱讀背面之注意事項再填寫本f ) 本纸張尺度適用中国國家標準（CNS ) A4規格（210X297公釐） 經濟部中夬樣準局員工消f合作社印製 A7 B7 五、發明説明（5 ) 一種以影像處理法製作的追蹤式立體顯像系統及方法， 其係藉由面孔探測技術與相關性的對比技術，發展出的 一套快速的觀察者追蹤我廣，藉以達成可線上操作的追 蹤式立體顯像系統。 本發明之以影像處理法製作的追蹤式立體顯像系統 所採用的面孔偵測技術係先由取像裝置攝取影像，形成 具有25個樣本的資料庫，再經由數據前處理，亦即先將 每個面孔中兩眼的位置調整到水平位置，再將每個面孔 由眉至上唇的區域，取成144(9 Ί 6)個晝素的畫面，以此 144個晝素所組成的向量為點，對資料庫的協方差矩陣 (Coyaria导ce matrix)作声有向量分析： €οϊ = Σ \xi - = ΣΛΣγ i 八=^Kw"-^s) Σ=[的··…. 每個固有向量a代表像空間中的一個基底方向，而 主要特徵向量〜則為此資料庫中最密集的方向轴，在先 則技術中使用固有向量作為偵測的方法，都是選擇數個 固有向量來組成低維空間’再以測試影像投影於此低維 空間所得到的座標，計算其可能性函數值，來判定影像 是否為面孔’基於群集背景的實際需求，以及追蹤系統 只追縱最主要之正面觀察者的考量，本發明運用兩段偵 測的方式’修改特徵臉的技術，增加偵測面孔可能存在 區域的階段，藉以解決群集背景的困擾，同時縮小找尋 局部相關最大值的區間1。另外，可提昇辨識率，以及大 姆尺度適用中國國 —Μ ----1 - ---批衣--I I K -- n 7 - (請先閲讀背面之注意事項再填寫本頁) 經濟部中央標準局員工消費合作社印製 41S793 A7 ------ B7 -___ 五、發明説明（6) 幅增加對面孔大小變化及歪斜的容忍度，而可進一步省 略各種比例的影像次取樣及重新整理的工作，可有效地 減少計算量。 為了更進一步地說明本發明之方法、架構及特徵， 姐配合附圖說明較佳實施例如下： 第1圖係燴示本發明之以影像處理法製作的追蹤式 立體顯像系統的架構圖式κ 第2圖係繪示本發明使用之面孔特徵模板的圖式。 第3圏係繪示本發明之以影像處理法製作的追蹤式 立體顯像方法之流程圖。 第4圖係繪示在本發明中追蹤觀察者面孔移動的方 法。， 實施例說呀 本發明之以影像處理法製作的追蹤式立體顯像系 統，包括：一成像裝置，用以相對於一觀察者而產生一 影像；一儲存裝置’儲存有一面孔特徵模板；一檢測裝 置，用以檢測在前述影像中觀察者影像的位置；一相關 裝置，用以對於連續輸入的觀察者的面孔影像進行相關 運算，以決疋觀察者的移動位置；一輸出裝置，用以將 觀察者的位置連續地輸出至立體顯像裝置$ 其中’前述檢測裝置係包括：一相關裝置，用以對 輸入的面孔影像與前述特徵模板進行相關運算；一第一 閾運算裝置，用以進行閾運算，藉以求出輸入面孔影像 經相關運算後，其較小的灰階值之分布區域；一過濾裝 本紙張尺度適用中国國家標隼（CNS ) A4規格（210x297公慶） HHV ^^^^1 n^— 4—^^ ^ϋϋ ^11 n n^v ^ ▼ » * - l - {請先閱讀背面之注意事項再填寫本頁) A7 413793 __B7 五、發明説明（7) 置，用以將前述經閾運算的影像加以過濾；一第二閾運 算裝置’用以對於過濾後的影像再次進行閾運算，藉以 決定面孔可能位置的區域；一局部最大值尋找裝置，用 以在上述面孔可能位置的區域中，找出相關性最大的位 置。 前面所描述的裝置之架構，可針對架構中各裴置的 功能’使用適當的電子電路元件來加以實施，然而在本 發明之實施例中’可直接使用一主電腦裝置，例如個人 電腦’來達成前述架構中之儲存裝置、檢測裝置、相關 裝置及輸出裝置的功能。 請參閱第1圖，本發明之以影像處理法製作的追蹤 式立艟顯像系統係包括：一攝影裝置10 ，例如一 CCD 攝影機，用以對一觀察者攝入一影像；一主電腦裝置2〇 , 例如一個人電腦’在該主電腦裝置20中儲存有一面孔特 徵模板，當接收到前述攝影裝置10輸入的影像時，前述 主電腦裝置20會將輸入的觀察者面孔影像與前述特徵模 板進行相關運算（correlationjj，然後先將前述相關運算的 結果進行閾運算（threshold)，藉以求出輸入面孔影像經相 關運算後’其較小的灰階值之分布區域，再將前述經閾 運算的影像加以過濾’接著對於過濾後的影像再次進行閾 運算，藉以決定面孔可能位置的區埵^同時並在上述面孔 可能位置的區域中’找出相關性最大的位置；一立體顯像 裝置30 ’以前述相關性最大的位置作為追縱的目標，並 據以分別投射立體顯像信號至觀察者50的兩眼中。 本紙張尺度適用中國國家標準（CNS ) A4規格（210X 29?公釐） J I -裝 — 訂 . t - (請先Μ讀背面之注意事項再填寫本頁) 經濟部中央榡準局負工消費合作社印製 413793 A7 B7 五、發明説明（8 ) ' ' 上述裝置更包括有一驅動裝置40，例如推挽式驅動 裝置’用以受追蹤到的面孔位置靜止或移動的影響，而 驅動一投射物鏡45 ’藉以使得由立體顯像裝置3〇投射 出的立體顯像信號始終保持被投射至觀察者5〇的兩眼 中。 _ .. ··. 前述主電腦裝置20中儲存的^^徵模板係預先經 計算求出，亦即先由攝影裝置10攝取一面孔影像，形成 具有25個樣本的資料庫’再經由數據前處理，先將每個 面孔中兩眼的位置調整到水平位置’再將每個面孔由眉 至上唇的區域，取成144(9x16)個畫素的畫面，以此144 個畫素所組成的向量/^點’對資料庫的協方差矩陣作固 i ^ = [ei^.....%] 經濟部中央樣準局員工消費合作社印裳 1^1 m vfn 士 ^^^1 m mi 1 - t' - 、-=# -. * · . {請先閱讀背面之注意事項再填寫本頁) 每個固有向量&代表像空間中的一個基底方向’而 主要特徵向量A則為此資料庫中最密集的方向軸。基於 群集背景的實際需求，及追蹤系統只追蹤最主要之正面 •觀察者的考量，本發明採取兩段式偵測系統，以此主要 特徵向量q作為特徵模板，如第2圖所未。 請參閱第3圖，其繪示以影像處理法製作的追蹤式 立體顯像方法之流程圖’包括下列步驟：步驟1〇〇， 利用前述攝影裝置拍攝觀察者面孔影像，輸入至主電腦 襞置，並利用主電腦裝置對輪入之面孔影像與特徵模板 本紙張尺度適用中_家標率（CNS ) A4規格（別χ 297公 413793 A7 -------- -B7 五、發明説明（9 ) ~- 進行相關私.；⑺步驟IG2，利用主電腦裝置先將前述 相關運算的結果進行閾運算，藉以求出輸入面孔影像經 相關運算後，其較小的灰階值之分布區滅⑺步驟1〇4， 將前述經閾運算的影像加以過濾；（4)步驟1〇6，利用主 電腦裝置對於過濾後的影像再次進行閾運算（5)步驟 108，將此閾運算的結果與步驟1〇〇之相關運算的結果進 行内積（inner pr〇duct)，藉以決定面孔可能位置的區域； (6)步驟110，在上述面孔可能位置的區域中，找出相關 性最大的位置，在此位置中的影像即為影像中正面主要 觀察者的面孔，可藉以作為立體顯像系統追蹤的目標。 在前述方法中，本發明修改了判斷特徵臉的技術， 増加貞測面孔可能存在區域的階段，可以解決群集背 景的困擾「，也可以減少所必需尋找的相關性局部極大值 之區間。 經濟部中央榡準局員工消費合作杜印製 --------•裝，------訂 - > f請先聞讀背面之注意事項再填寫本頁) 此外，運用前述本發明之方法可提昇辨識率，也可 大幅增加對面孔大小變化及歪斜的容忍度，故而可進一 步省略各種比例的影像次取樣與匹配的操作，可以有效 地降低所需的計算量’所以在進行追蹤面孔影像時，可 減少影像處理運算所需的時間。 請參閱第4圖，在本發明之以影像處理法製作的追 蹤式立體顯像方法中，更包括下列步驟（丨）對於連續輸 入的觀察者的面孔影像，亦即已偵測到的面孔影像DI與 緊接著輸入的影像INPUT進行相關運算，以決定觀察者 移動後的位置CL ; (2)根據觀察者的位置CL資料，主電 12 本纸張尺度適用中g|國家標翠（CNS ) A4規格（：丨O xiW公趁) ~— 413793 A 7 B7____ 五、發明説明（10 ) 腦裝置可控制驅動裝置，藉以使立體顯像裝置可以輸出 影像訊號至觀察者面孔上的兩眼t。上述之步驟即在於 偵測觀察者面孔的移動，透過連績輸入的兩個面孔影像 的相關運算，簡易地求出觀察者移動後的面孔位置。 (請先閱讀背面之注意事項再填寫本頁) 經濟部中央橾牟局貝工消費合作社印製 本紙張尺度適用中圉國家橾準（CNS ) A4規格（210X 297公釐）Regarding the tracked stereoscopic imaging system and method V produced by the image processing method in the development process of virtual reality technology_, in order to allow the observer to view the stereoscopic image, the observer must generally be required to carry a special viewing Devices, such as glasses, to transmit different images to the eyes of the observer. However, arranging these viewing devices will cause restraint on the observer, and the movement range of the observer is also limited. In order to improve such disadvantages, in the prior art, a tracking stereoscopic imaging system has also been developed. By using the information of the observer's location, it is possible to reduce the concentration, There are currently five methods of tracking system applied in virtual reality: (1) Mechanical type: it is composed of a link device. The device using this method will restrict the user and the range of movement. It will be similarly limited. "(2) The electromagnetic type" uses three receiving antennas to receive three consecutive sets of electromagnetic waves, "using nine electromagnetic field strength data to calculate the position and direction of the user's head, but the delay time Long, tracking range is small, not suitable for the situation where the user is constantly moving, and easily affected by the external magnetic field β (3) Ultrasonic induction type, its principle is similar to the electromagnetic type, that is, the difference between the transmission time of three groups of ultrasonic waves is used to Triangular positioning, but its accuracy is easily affected by changes in air density or obscured by other objects. (4) Optical type, which uses laser or infrared light as the light source, and is installed in the surrounding environment in a specific arrangement. At the same time, the head-mounted display has a sense of setting. 1G X plus public $) —'— ----— ~-Packing ------ Order --'--- r (Please read the notes on the back before filling this page) 413793 Central Bureau of Standards, Ministry of Economic Affairs Consumption cooperation with employees Du printed A 7 B7 5. Description of the invention (2) The sensor uses the current generated by the photocell and the position of the light center on the sensor to estimate the position and direction of the head, but its tracking range is affected by Limitation. In the prior art, in order to provide users with the most natural stereoscopic imaging method, image processing technology using face detection has been developed to determine the position of the observer's face, which can be used as a basis for tracking the position of the eyes of the observer to guide The stereoscopic imaging system focuses and projects the image into the viewer's eyes. The previous technologies related to such image processing-type tracking stereoscopic imaging systems and methods can be roughly divided into the following methods: 1. Neural network-based, such as (l) SH Lin et al. "Face Recognition / Detection by Probabilistic Decision-Based Neural Network" published in Wewra / Vol_ 8, pp. 114-132, Jan. 1997, which adopts two-stage learning correction, first is local unsupervised learning, each subnet Training separately, using the EM method to estimate the subspace parameters composed of high-frequency components in the image. After this regional training is completed, it enters globally supervised training, and then the boundaries obtained from the regional training are further refined based on the data. Adjustment. (2) In "Convolutional Neural-Network for Face Recognition" of S. Lawrence et al., Proc. IEEE Conf.on Computer Vision Sin £ ^ corporate < 3 " ^ «and 6 (^ 尽 则 '" 〇 «, 卩 卩 .217-: 222, 1996_, which combines self-organizing map network and convolutional neural network. Self-organizing map neural network can ( Please read the notes on the back before filling in this I) This paper size is applicable to the Chinese National Standard (CNS) A4 specification (210X297 mm) 413793 a7 B7 Printed by the Consumers ’Cooperative of the Central Government Bureau of the Ministry of Economic Affairs 5. Description of the invention (3) Reduce the input dimension (the gray scale with a resolution of 92'112), so that the image is quantified in topological space, while the convolutional neural network provides a part of the displacement, rotation, size change, and deformation. Invariance, the recognition rate in this report is better than the intrinsic vector method. (3) In " NeuralNetwork-BasedFaceDetection ', by HA Rowley et al., Proc. IEEE Conf. On Computer Vision and Pattern Recognition, 1996. First Illumination and histograms are converted into data pre-processing, and then possible windows are detected with various window sizes and sample ratios. This method uses a neural network trained with 1050 trained face images and 130 tested faces. It can reach a recognition rate of 90.5 ° /. 2. Feature-based face recognition is first assembled on the facial features by a variable broadcast template (template), the geometric shapes of the facial features are extracted, and then normalized. (Normalize) into a normal form, normalized appearance-'^ The image is then classified by a set of main thinking vectors. The representative vectors used in this classification can be compared to the five features without being affected by the complex influence of expression changes. In IJ Cox et al's "Feature-based Face Recognition Using Mixture-Distance", Proc. IEEE Conf. On Computer Paiierw1996 Yin, with six potential energy items including the relationship between variables, in a cycle of repeated template assembly process , Each parameter is gradually modified, but the recognition rate obtained by this method is not ideal. Therefore, in MAShackleton et al's "Classification of Facial Features for Recognition", Proc. IEEE Conf. On 1991, a mixture-distance technique is used, each face consists of 35 quantities (please read first Note on the back, please fill in this page again) Packing · 111 This paper size is applicable to China National Standard (CNS) A4 (210X297 mm) Printed by the Consumers Cooperative of the Central Standards Bureau of the Ministry of Economic Affairs A 7 B7 V. Invention Description (4) The 30 distances estimated from the measured positions are representative, and can reach a recognition rate of 95%, but the calculation amount is too large. Third, the visual-based feature space, this method assumes that all possible face patterns occupy a small and variable determined subspace. This method uses the data group and the main component in one or more sample groups of the face image to approximate the subspace of the face pattern. \ The feature space approach has the advantage that it can handle closed situations. However, this approach has only proven to work in a non-cluster context. Related materials can be found in 1'Face Recognition Using Eigenfaces ", Proc. IEEE Conf. On Computer Vision and Pattern Recognition, 1991, T. Darrell et al. &Quot; Active Face Tracking and Pose Estimation in an MA Turk et al. Interactive Room ", Proc. IEEE Conf. On Computer Vision and Pattern Recognition, 1996, and U.S. Patent No. 5,1 64,992. This method uses eigenfaces for face detection, which is fast and simple. However, its detection depends on the high correlation between pixels in the training and test images. In a stereoscopic imaging system, it is necessary to track the face position of the main front observer in front of the system, and within an acceptable amount of calculation, continuously rotate the observer's dual-purpose position to serve as the target space of the stereoscopic imaging. However, in the aforementioned prior art, the system of face recognition and tracking has problems such as excessive calculation amount, large manual processing, and lengthy derivation and convergence of cyclic and iterative operations. In view of this, in order to reduce the repeated focus caused by time multiplexing and space multiplexing in a stereoscopic imaging system, the object of the present invention is to provide n Hu ^ 1. ^ 1 ^^ 1 · 1 I—I HI —I u-_ -1— I, T----(Please read the notes on the back before filling in this f) This paper size applies to China National Standard (CNS) A4 (210X297 mm) Ministry of Economic Affairs Sample A7 B7 printed by the staff of the Prospective Bureau Cooperative Cooperative V. Description of the invention (5) A tracking stereo imaging system and method made by image processing method, which is developed by the face detection technology and the correlation technology of correlation A fast set of observers followed me to achieve a tracking stereoscopic imaging system that can be operated online. The face detection technology used in the tracking stereoscopic imaging system produced by the image processing method of the present invention is to first capture images from the image capturing device to form a database with 25 samples, and then pre-process the data, that is, first The position of the two eyes in each face is adjusted to the horizontal position, and then the area from each eyebrow to the upper lip is taken to form a picture of 144 (9 Ί 6) celestial elements. The vector composed of 144 celestial elements is Point, the vector analysis of the database's covariance matrix (Coyaria ce matrix): € ο 有 = Σ \ xi-= ΣΛΣγ i eight = ^ Kw "-^ s) Σ = [的 ··…. Each The eigenvector a represents a base direction in the image space, and the main feature vector ~ is the most dense direction axis in the database. In the prior art, the eigenvector is used as the detection method. Several eigenvectors are selected. To form a low-dimensional space, and then use the coordinates obtained by projecting the test image onto this low-dimensional space to calculate its likelihood function value to determine whether the image is a face. Based on the actual needs of the cluster background, and the tracking system only pursues the most important Positive view Those considerations, the present invention is the use of two-stage embodiment of Detection 'eigenfaces modification techniques, there may be an increase phase in the face detection area, so as to solve the problems of the background cluster while reducing an interval to find local correlation maxima. In addition, the recognition rate can be improved, and the size of the country is applicable to China—M ---- 1---- Approval --IIK-n 7-(Please read the precautions on the back before filling this page) Ministry of Economic Affairs Printed by the Consumer Standards Cooperative of the Central Bureau of Standards 41S793 A7 ------ B7 -___ V. Description of the invention (6) Increased tolerance for changes in face size and distortion, and can further omit sub-sampling and re-image of various proportions of images Finishing work can effectively reduce the amount of calculation. In order to further explain the method, architecture, and characteristics of the present invention, the following describes a preferred embodiment with reference to the accompanying drawings: FIG. 1 is a diagram showing the architecture of a tracking type stereoscopic imaging system made by image processing of the present invention κ Figure 2 is a diagram showing a facial feature template used in the present invention. The third aspect is a flowchart of the tracking type stereoscopic imaging method made by the image processing method of the present invention. Fig. 4 illustrates a method for tracking the movement of the face of an observer in the present invention. According to the embodiment, the tracking type stereoscopic imaging system made by the image processing method of the present invention includes: an imaging device for generating an image relative to an observer; a storage device 'storing a facial feature template; A detection device is used to detect the position of the observer's image in the aforementioned image; a correlation device is used to perform correlation operations on the face image of the observer continuously input to determine the position of the observer's movement; an output device is used to The position of the observer is continuously output to the stereoscopic imaging device. 'The aforementioned detection device includes: a correlation device for performing a correlation operation on the input face image and the aforementioned feature template; and a first threshold operation device for Threshold calculation is performed to find the distribution area of the smaller grayscale values of the input face image after correlation calculation; a filter paper size is applicable to China National Standard (CNS) A4 size (210x297 public celebration) HHV ^^ ^^ 1 n ^ — 4 — ^^ ^ ϋϋ ^ 11 nn ^ v ^ ▼ »*-l-(Please read the notes on the back before filling this page) A7 413793 __B7 V. Send Explanation (7) is configured to filter the aforementioned thresholded image; a second threshold computing device is used to perform threshold computation on the filtered image again to determine the area where the face may be located; a local maximum search A device for finding the position with the highest correlation among the possible positions of the face. The architecture of the device described above can be implemented using appropriate electronic circuit components for each function in the architecture. However, in the embodiment of the present invention, a host computer device such as a personal computer can be used directly. The functions of the storage device, the detection device, the related device, and the output device in the foregoing architecture are achieved. Please refer to FIG. 1. The tracking type display system produced by the image processing method of the present invention includes: a photographing device 10, such as a CCD camera, for ingesting an image to an observer; a host computer device 20. For example, a personal computer 'stores a facial feature template in the host computer device 20. When receiving an image input by the aforementioned photographing device 10, the host computer device 20 performs an input of the observer's face image with the aforementioned feature template. Correlation operation (correlationjj, and then threshold the result of the aforementioned correlation operation to obtain the distribution area of the input gray image's smaller grayscale value after the correlation operation, and then the aforementioned threshold image Filter it 'and then perform a threshold operation on the filtered image again to determine the possible positions of the face 同时 ^ At the same time,' find the most relevant position in the area of the possible positions of the face; a stereoscopic imaging device 30 'to The aforementioned most relevant positions are used as tracking targets, and the three-dimensional imaging signals are respectively projected to the observer 50. In the eyes. This paper size is in accordance with Chinese National Standard (CNS) A4 (210X 29? Mm). JI-binding-binding. T-(Please read the notes on the back before filling this page) Printed by the Industrial and Consumer Cooperatives 413793 A7 B7 V. Description of the invention (8) 'The above device further includes a driving device 40, such as a push-pull type driving device, which is used to be affected by the position of the tracked face to be stationary or moving, and drive a The projection objective lens 45 ′ keeps the stereoscopic imaging signals projected by the stereoscopic imaging device 30 from being constantly projected into both eyes of the observer 50. _ ..... Stored in the aforementioned main computer device 20 The template is obtained through calculation in advance, that is, a face image is captured by the photographing device 10 to form a database with 25 samples 'and then the position of the two eyes in each face is adjusted to a horizontal position through data pre-processing' Then take the area of each face from the eyebrow to the upper lip to take a picture of 144 (9x16) pixels, and use the vector / ^ point 'composed of 144 pixels to fix the covariance matrix of the database. ^ = [ei ^ .....%] in the Ministry of Economic Affairs Sample quasi bureau employee consumer cooperatives Yin Chang 1 ^ 1 m vfn ^^^ 1 m mi 1-t '-,-= #-. * ·. (Please read the precautions on the back before filling this page) Each inherent The vector & represents a base direction in image space 'and the main feature vector A is the densest direction axis in this database. Based on the actual needs of the cluster background, and the tracking system only tracks the most important fronts. • For the observer's consideration, the present invention adopts a two-stage detection system, using the main feature vector q as a feature template, as shown in Figure 2. Please refer to FIG. 3, which shows a flowchart of a tracking stereoscopic imaging method produced by image processing method, including the following steps: Step 100: Use the aforementioned photographing device to capture an image of an observer's face, and input it to the host computer. ， And use the main computer device to rotate the face image and feature template. This paper scale is applicable in the _ house standard rate (CNS) A4 specification (not χ 297 public 413793 A7 -------- -B7 V. Description of the invention (9) ~-Perform correlation. 相关 Step IG2, using the main computer device to threshold the result of the aforementioned correlation operation to obtain the smaller grayscale value distribution area of the input face image after the correlation operation. Step 104 is eliminated, and the aforementioned threshold image is filtered; (4) Step 10, the host computer device is used to perform threshold calculation on the filtered image again; (5) step 108, the result of this threshold operation An inner product is performed with the result of the correlation operation in step 100 to determine the area where the face may be located; (6) Step 110 is to find the position with the highest correlation among the areas where the face may be located, in The image in the position is the face of the main observer in the image, which can be used as the target for tracking by the stereoscopic imaging system. In the foregoing method, the present invention modifies the technique for determining the characteristic face, and increases the stage of measuring the possible existence of the face It can solve the problem of cluster background ", and it can also reduce the interval of the local maxima of the correlation that must be found. The consumer cooperation of the Central Bureau of Standards of the Ministry of Economic Affairs is printed -------- • equipment, --- --- Order- > f Please read the notes on the back before filling in this page) In addition, using the method of the present invention can improve the recognition rate, and can also greatly increase the tolerance for face size changes and skew. Further omitting the operations of sub-sampling and matching of images in various proportions can effectively reduce the amount of calculation required ', so when tracking face images, the time required for image processing operations can be reduced. Please refer to FIG. 4, in the present invention The tracking stereoscopic imaging method produced by the image processing method further includes the following steps (丨) for the continuous input of the observer's face image, that is, the detected The face image DI obtained is correlated with the input image INPUT to determine the position CL of the observer after moving; (2) According to the data of the observer ’s position CL, the main paper 12 is applicable to g | National Standard Cui (CNS) A4 specification (: 丨 O xiW) while taking advantage of ~~ 413793 A 7 B7____ 5. Description of the invention (10) The brain device can control the driving device, so that the stereoscopic imaging device can output image signals to the observer's face Both eyes t. The above steps are to detect the movement of the observer ’s face, and to easily find the position of the face after the observer ’s movement through the correlation calculation of the two face images input by the succession. (Please read the precautions on the back first (Fill in this page again.) Printed by the Central Ministry of Economic Affairs of the Ministry of Economic Affairs, Shellfish Consumer Cooperatives. This paper is printed in accordance with the Chinese National Standard (CNS) A4 (210X 297 mm).