200919382 九、發明說明： 【l月所屬之技術領域】 t 明传 ,,,t ’、—種有關於跌倒偵測照護系統之技術領域， 尤才日一種運算 照護系統 丈有政率的具多格影像處理能力之跌倒偵測 【先前技術】 ^ n 、跌倒偵測照護系統—般係請使用者佩帶加速器 导微感測儀考於包L ^ .A ,, ° ;身上，再利用人活動之訊號進行辨識，易 造成使用者3崔他 刃 判。其欠:可 便，且沒有現場晝面輔助容易造成誤 主ν α。人習用利用攝影機進行跌倒债測照護的系統，多 1單—攝影機的影德声柿边+ 機，且未對多師像串:：若直接擴充為多台攝影 像串〜έ又計加速計算演算法，則系統效 月匕與衫像品質將非常低落。 再者，目前市面上常見的 a 私 視糸統，疋利用硬體 擷取卡外接多台攝影機進 弋銪„。丨 —甶瓜控，但僅有影像傳送， 或間早偵測異物入侵之功能， 、 5, ^ ^ 未犯進仃物件識別功能，以 致…法應用於跌倒偵測照護相 昭罐奉轉柄户卜" 又相關之服務，習用的跌倒偵測 …系統仍存在诸多的限制與 【發明内容】 嗵手-有改良之必要。 欲解決之技術問題點··習 ..^ 用的跌倒偵測照護系統—护 係。“吏用者佩帶加速器等微感測儀器於身上 又 動之訊號進行辨識，易造成使用者攜帶 活 場畫面輔助容易造成誤判。其次，羽 ，又有現 倒偵測照護的系統，多半以單1用攝影機進行跌 早攝影機的影像處理為主， 5 200919382 若直接擴充為多a ά 計算渖曾法 Γ办’’且未對多格影像串流設計加速 丨介6、开去’則系統 目前市面上常見的保全&視二品f將非常低落。再者， 多台攝影機進行書：“，是利用硬體操取卡外接 異物入侵之功僅有影像傳送，或簡單伯測 於跌倒價測照護相^ 2件識別功以致無法應用 李统仍存在#服務等問題’習用的跌倒偵測照護 糸、.先仍存在啫多的限制與問題等。 2問題的技術特點：提供一種具多 :=測照護系統，係包括I複數台攝影機、-多格影 與至少—保全端警示裝置。其中，各攝影機 網路與該多格影像處理― 記憶體結構傻/多格影像處判服S包括一 …〜格4處理伺服器係、可透過從各攝影機 :4影像串流資料進行影像處理與跌倒樣型識別演 別是否有跌倒意外發生，該影像處理與跌倒樣型 8〜^法係包括四個執行緒： 义I影像操取執行緒，該影像掏取執行緒係為系統啟動 :壯無人場景圖’供之後比對用，當該攝影處 王衣置啟動後’該攝影機係以固定速率拍攝現場畫面，並 將現％畫面儲存至該記憶體結構内； b.影像處理執行緒，該影像處理執行緒係將現場畫面 中與原本無人場景圖之間差異的物體’利用影像處理的動 作過渡出來’使得該物體影像不會失真； C•人形輪廓萃取執行緒，該人形輪廓萃取執行緒係將 200919382 影像中的該物體的輪廓描繪出來，做為人形判斷. d.跌倒樣型識別執行緒，該跌倒樣型識別執行緒係為 將現場晝面的該人形輪靡與多張跌倒的樣型作比對，該跌 倒樣型係預先取出的輪廓特徵點並儲存於系統中，因此現 場拍攝畫面的人形輪廓特徵與該跌倒樣型 比對後，即可根據比對結果來判讀跌倒意外是否\=丁逐 該多格影像處理伺服器係可進行多格影像串流運曾， 該f格影像串流運算為多執行緒（Multithreading)與^線 平行⑺peUning)運算，各執行緒負責單—步驟的工作， 所有執行绪係依據管線化排程運作，藉此，當第一台攝影 機的第-影像串流資料進人系統後，該影錢取執㈣= 執行影像擷取之運算，待運算完成後，操取的影像會被儲 存至該記憶體結構内，接著由該影像處理執行緒進行影像 處理動作’同—時間，該影像擷取執行緒立即接收第二a 攝影機的第二影像争流資料，重覆影像擷取工作，待 像擷取執行緒與該影像處理執行緒平行運算後，該第」: 像串流資料輪由該人形輪廓萃取執行緒運算，該第二影= =流貢料交由該影像處理執行緒運算，以此類推，系統最 ，可同時存在四個平行運算之執行緒，藉此可達同—時間 多工處理，藉以加速影像運算之功效者；該保全端警示裝 置係包括一顯示螢幕’該保全端警示裝置係與該多格影像 處理伺服器可彼此進行通訊地連接； 藉此，當社區人員發生跌倒意外時，該攝影機擷取該 衫像串流貧料傳送至該多格影像處理伺服器進行該影像處 200919382 理與跌倒樣型識別演算法谋 f异做出判斷 其中 其中 其中 其中 其中 保全端警示裝置的該顯示螢幕上·』辦，並發送警訊到該 警訊後，可進行即時救護與保八動 s彳軒或醫護人員收到 其中，該保全端警示裝 作2 該手持式設備可為—Pda ^備。 該手持式設備可為—手機/ 該手持式設備可為— 為旱上型電腦。 该保全端警示裝置可 該網路可為“ 4纟上型電腦。 有、、泉網路或無線網路 對照先前技術之功效： 本發明係利用寻;；@ + Tm t 、’处里14樣型識別技術 照護之服務，使用本a y 凡成跌倒偵測 ^ m 热須佩帶任何微感測器於身上。 本务明開發出跌倒影像辨識演算 拍攝不同現場金& 扪用夕台攝影機 跌倒”之樣型=匕’Γ掏取人形輪廣特徵點並與，，類 a 進仃比對，可大幅簡化影像處理計算量， 且不會影響辨識正確率或誤判率。 開發出多格影像串流加速計算演算法，可大幅 文α糸統影像處理效能。 四 本發明在辨1山1 4出可能發生人員跌倒事件時，即時傳送 詈告簡訊给盤芦-_J_、 五 、s知或醫護人員前往察看。 本毛明知用個人化電腦標準架構開發，無須另外附加 特定之軟體或硬體模組。 有關本發明印_ p m ^ ^ 所知用之技術、手段及其功效，茲舉一較 4實施例並配合円、， σ圃式砰細說明如后，相信本發明上述之目 200919382 的、構造及其特徵’當可由之得—深入而具體的 【實施方式】 請參閱第一圖至第七圖所示，本發明係提供-種具多 格影像處理能力之跌倒備測照護系統，係包括：複數台攝 影機（1 〇 )、一多格影像處理伺服器（2 0 )盘至少一 保全端警示裝置（30)。其中，各攝影機（1〇)係用 乂心貝取如像串机貝料，該多格影像處理飼服器（2 )係 透過網路與各攝影機（1〇)進行連接，該多格影像處理 伺服时（2 0 )包括一記憶體結構，該多格影像處理词服 =(2 0 )係可透過從各攝影機"〇 )擷取來的該影像 串流資料進行影像處理與跌倒樣型識別演算法來判別是否 有跌倒意外發吐，兮旦〈你士油μ .一 忒心像處理與跌倒樣型識別演算法係包 括四個執行緒： a ::擷取執行緒’該影像擷取執行緒係為系 :攝先广張無人場景圖（6〇)，供之後比對 拍攝現場書面影·（1〇)係以固定速率 記憶體結構内），亚將現場畫“7。)儲存至該 (乂：理執行緒’該影像處理執行緒係將現場晝面 ’、原本無人場景圖（6 0 )之間差異的物體（8 真；，用影像處理的動作過遽出來，使得該物體不會失 將 •人形輪料取執行緒，該人形輪料取執行緒係 200919382 的輪廓 0)描繪出來，做為 影像中的該物體（8 〇 人形判斷； d.跌倒樣型識別執行緒，該 將頦尸蚩品r π n』像坦識別執行緒係為 將現场晝® ( 7 〇 )的該人形 巧 楛荆从L u v y ◦)與多張跌倒的 樣型作比對，該跌倒樣型係預先取 ^ έ ^ . 山曰7輪琛特徵點並儲存 於乐統中，因此現場拍攝畫面的人 aiL 柄々卩（9 0 )特徵盥 3亥跌倒樣型的特徵進行逐一比對後， 了俊即可根據比對結果來 判5胃跌倒意外是否發生； 士該多格影像處理飼服器（2〇)係可進行多格影像串 :運异，該多格影像串流運算為多執行緒（Mimithreading) 、管線平行（p1Pellning)運算，各執行緒負責單一步驟的 2作，所有執行緒係依據管線化排程運作，藉此，當第一 /攝々機的第-影像串流資料進入系統後，該影像掏取執 2緒即執行影像擷取之運算’待運算完成後，搁取的影像 2儲存至該記憶體結構内，接著由該影像處理執行緒進 y影像處理動作，同一時間，該影像擷取執行緒立即接收 第^台攝影機的第二影像串流資料，重覆影像擷取工作， 待忒影像擷取執行緒與該影像處理執行緒平行運算後，該 影像串流資料輪由該人形輪廓萃取執行緒，該第二影 像串流資料交由該影像處理執行緒運算，以此類推，系統 最古200919382 IX. Description of invention: [Technical field to which l month belongs] t Ming Chuan,,, t ', - is a kind of technical field about fall detection and care system, especially a kind of computing care system Fall detection of image processing capability [Prior Art] ^ n, fall detection and care system - the user is required to wear the accelerator to guide the micro-sensor to test the package L ^ .A,, °; body, reuse people activities The signal is identified, which is easy for the user 3 to judge. It owes: it is easy, and there is no on-site squatting aid that is easy to cause the wrong ν α. People use the camera to carry out the system of falling debt measurement and care, more than one single - the camera's phantom sound persimmon + machine, and not for the multi-teacher string:: If directly expanded into multiple photographic strings ~ έ Acceleration calculation The algorithm, the system efficiency and the quality of the shirt will be very low. In addition, the current private video system on the market, the use of hardware to pick up the card to connect multiple cameras into the 弋铕. 丨 甶 甶 甶 ,, but only the image transmission, or early detection of foreign body invasion Function, 5, ^ ^ No object recognition function, so that the method is applied to the fall detection and care, and the related services, the conventional fall detection... the system still has many Restrictions and [Invention Contents] 嗵手-There is a need for improvement. The technical problem to be solved.················································································· The identification of the moving signal is easy to cause the user to carry the live screen assist and easily cause misjudgment. Secondly, Yu, there is a system for detecting and repairing the current situation. Most of them use the camera to perform image processing of the early fall camera. 5 200919382 If it is directly expanded to more than one ά, the calculation is done. Multi-image video streaming design speeds up the introduction, and then the system is currently on the market, the common security & second product f will be very low. In addition, a number of cameras carry out the book: "It is the use of hard gymnastics to pick up cards for external foreign body invasion. Only the image transmission, or simple measurement of the fall price measurement and protection phase 2 2 identification work can not be applied to Li Tong still exist # Services and other issues 'practical fall detection and care 糸,. There are still many limitations and problems. 2 Technical characteristics of the problem: provide a multi-purpose: = measurement and care system, including I complex camera, - multi-grid Shadow and at least - security warning device. Among them, each camera network and the multi-image processing - memory structure silly / multi-image image judgment S includes a ... ~ grid 4 processing server system, can be transmitted from each camera : 4 image streaming data for image processing and fall type recognition performance whether there is a fall accident, the image processing and fall sample 8 ~ ^ system includes four threads: Yi I image manipulation thread, the image The thread is triggered by the system: the unmanned scene map is used for comparison. When the camera is opened, the camera shoots the scene at a fixed rate and saves the current % screen to the Within the memory structure; b. image processing thread, the image processing thread is to transfer the object 'distance from the image processing action' between the original scene and the original unmanned scene image to make the object image not distorted; • Humanoid contour extraction thread, which depicts the outline of the object in the 200919382 image as a humanoid judgment. d. Falling the sample recognition thread, the fall type recognition thread is The humanoid rim on the surface of the scene is compared with a plurality of falling patterns, which are pre-extracted contour feature points and stored in the system, so that the human contour feature of the live shooting picture is compared with the falling shape After that, it is possible to judge whether the fall accident is based on the comparison result. If the multi-image image processing server system can perform multi-image streaming, the f-image stream operation is multithreading. Parallel with the ^ line (7) peUning) operation, each thread is responsible for the single-step work, all threads are operated according to the pipelined schedule, thereby, when the first camera After the first video stream data enters the system, the money is taken (4) = the image capturing operation is performed, and after the operation is completed, the processed image is stored in the memory structure, and then processed by the image. The thread performs the image processing action 'same-time, the image capture thread immediately receives the second image contention data of the second a camera, repeats the image capturing work, and waits for the image capturing thread and the image processing thread After the parallel operation, the first: the stream data wheel is extracted by the humanoid contour extraction thread, the second shadow == stream tribute to the image processing thread operation, and so on, the system is the most, can exist at the same time The four parallel computing threads can be used to achieve the same-time multiplexing processing, thereby speeding up the function of the image computing; the security warning device includes a display screen, the security warning device and the multi-image processing The servers can be communicatively coupled to each other; thereby, when a community member has a fall accident, the camera captures the shirt and sends the image to the multi-image processing server for the image. At 200919382, the fall and fall-type recognition algorithm is used to make judgments, among which the display screen of the warning device is installed, and the alarm is sent to the alarm, and immediate rescue and protection can be performed. Eight mobile s彳 Xuan or medical staff received it, the security terminal warning pretending to be 2 The handheld device can be -Pda. The handheld device can be - the mobile phone / the handheld device can be - a dry type computer. The security warning device can be a "4" type computer. The utility model has the function of the prior art: the invention utilizes seek;; + + Tm t, 'where 14 Sample identification technology care service, use this ay Fan Cheng fall detection ^ m heat must wear any micro sensor on the body. Ben Ming developed a fall image recognition algorithm to shoot different scenes of gold & 扪 夕台相机 falls "The sample type = 匕" captures the wide feature points of the humanoid wheel and compares with the class a, which can greatly simplify the image processing calculation without affecting the recognition accuracy rate or false positive rate. Developed a multi-image image stream acceleration calculation algorithm, which can greatly improve image processing performance. 4. The invention sends an obituary message to Panlu-_J_, V, S, or medical personnel to see when a person may fall into the event. Ben Maoming is developed using a personal computer standard architecture without the need to attach specific software or hardware modules. Regarding the techniques, means, and effects of the present invention, it is to be understood that the techniques, means, and effects of the present invention are the same as those of the fourth embodiment, and the σ 圃 圃 如 , , 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 2009 And its characteristics 'when it can be obtained - deep and specific [implementation] Please refer to the first to seventh figures, the present invention provides a multi-image image processing capability of the fall preparation and care system, including : a plurality of cameras (1 〇), a multi-frame image processing server (20), at least one full-end warning device (30). Among them, each camera (1〇) is connected with a heart-shaped shell, such as a string machine, and the multi-image image processing feeder (2) is connected to each camera (1〇) through a network, and the multi-frame image processing is performed. The servo time (20) includes a memory structure, and the multi-image processing word service = (2 0) is capable of performing image processing and falling samples through the image stream data captured from each camera "〇) Identify the algorithm to determine if there is any accidental vomiting, and you will have a four-thread: a: 撷 执行 ' ' 撷 该 该 该 该 该 该 该 该 该 该 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌 跌The thread is the system: the first unmanned scene map (6〇), for later comparison, the written scene of the scene (1〇) is in the fixed rate memory structure), the Asian will be painted "7.) To the (乂: Logic Thread 'The image processing thread is the scene of the scene ' ', the original unmanned scene map (60) difference between the objects (8 true;, through the action of the image processing, so that The object will not be lost. • The humanoid wheel takes the thread and the humanoid wheel takes the execution. The outline of the 200919382 is depicted as the object in the image (8 〇 humanoid judgment; d. fall type recognition thread, the 颏 蚩 r r r 』 像 像 像 识别 识别 识别 识别 识别 识别 识别 识别 识别 识别 识别 识别 识别Field 昼® ( 7 〇 的 该 从 从 从 从 从 从 从 从 从 从 从 从 从 从 从 从 与 与 与 与 与 与 与 与 与 与 与 与 , , , , , , , , , , , . . . . . . . . . . . . . . In the music system, the characteristics of the person aiL 々卩(9 0) on the scene shooting screen are compared one by one, and then Jun can judge whether the stomach fall accident occurs according to the comparison result; The multi-image image processing feeding device (2〇) can perform multi-image image stringing: the multi-image video stream computing operation is multi-threaded (Mimithreading), pipeline parallel (p1Pellning) operation, each thread is responsible for In the single step, all the threads are operated according to the pipelined schedule, so that when the first video stream data of the first/camera enters the system, the image capture algorithm performs image capture. Operation 'after the completion of the operation, the image 2 that was put on is stored in the memory In the structure, the image processing operation is followed by the y image processing operation. At the same time, the image capturing thread immediately receives the second image stream data of the second camera, repeating the image capturing work, and the image to be scanned 撷After the thread is paralleled with the image processing thread, the image stream data wheel is extracted by the humanoid contour extraction thread, the second image stream data is subjected to the image processing thread operation, and so on, the system is the oldest
巧可同時存在四個平行運算之執行緒，藉此可達同一時 間 S 夕工處理，藉以加速影像運算之功效者；該保全端警示 、（3 0 )係包括一顯示螢幕，該保全端警示裝置（3 〇 )係與該多格影像處理伺服器（2 〇 )可彼此進行通訊 10 200919382 地連接； 藉此’當社區人員發生跌倒音 擷取該影像串流資料 夸邊攝影機（1 〇〕 貝7H·傳运至該多才夂吾 進行該影像處理與跌倒樣型㈣理賴器（2〇) 發送警訊到該保全端警示 =法運算做出判斷，並 該警衛（4 〇 )或醫ftt人。,U )的該顯示螢幕上； 即時救護與保全動作。貞5 Q )收到警訊後’可進行 其中，該保全端警示裝置（ 其中，兮手拄（ 3 〇 )可為一手持式設備。 其中 其中 其中 其中 此外 ' T °亥手持式設備可為—p D Α。 該手持式設備可為—手機。 該手持式設備可為—掌上型電腦。 该保全端警示裝置 〇)可為一桌上型電腦。 该肩路可為有線網路或無線網路。 本發明的實施例的詳細八 月刀述如下，本發明 的使用情境係如第一圖所示。 在第一圖中，假設老年人常 居動且易發生跌倒意外的場 一 琢π如人仃道、運動場、公園、 田場等，我們在各場所架設攝 砰〜栻（1 〇 )，並將影像 傳回多格影像處理伺服器（2 0 ^ ^ )中進仃跌倒樣型識別。 虽該多格影像處理伺服器（2 0) J y別畫面中出現跌倒意 外時’便立即傳送手機簡訊給鍪 σ -俯（4 0 )或醫護人員（5 0 )。 在多格影像處理伺服器（2 η ) μ ΑΑ么 V d U )上的糸統晝面，如第 圖所示。我們將多格晝面排列在金 J仕旦面上方，並依序編號 200919382 為〇號攝影機、Μ攝影機."等。畫面下方㈣出9種“類 似跌倒，，的樣型以供識別。在每一格影像旁邊皆列出目^ 晝面與跌倒模型比對之後的相似程度’數字愈小表示愈相 似。當連續拍攝的晝面都與同一跌倒樣型相似時，系統便 發出手機警告簡訊’畫面如第三圖所示。 夕、下刀別況明本系統單格影像的跌倒辨識演算法以及 多格影像串流加速運算演算法。 一 '跌倒影像辨識演算法： 跌倒衫像辨識共分為四個步驟，分別是： 步驟一：影像擷取執行緒It is possible to have four parallel computing threads at the same time, so as to achieve the same effect of the S time processing, thereby speeding up the effect of the image computing; the security warning, (3 0) system includes a display screen, the security end warning The device (3 〇) and the multi-image processing server (2 〇) can communicate with each other 10 200919382; thereby, when the community member has a fall sound, the video stream data quart camera (1 〇) Bell 7H·Transported to the versatile 夂 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行 进行Ftt person., U) of the display on the screen; instant rescue and security actions.贞5 Q) After receiving the warning, the security warning device (where the handcuffs (3 〇) can be a handheld device. Among them, the 'T °H handheld device can be —p D Α. The handheld device can be a mobile phone. The handheld device can be a palmtop computer. The security alert device can be a desktop computer. The shoulder can be wired or wireless. The detailed August of the embodiment of the present invention is as follows, and the usage scenario of the present invention is as shown in the first figure. In the first picture, we assume that the elderly are often active and prone to fall accidents, such as human squats, sports fields, parks, fields, etc., we set up photos in each place ~ 栻 (1 〇), and The image is sent back to the multi-image processing server (2 0 ^ ^ ) for the fall and fall type recognition. Even if the multi-image processing server (20) has a fall in the J y screen, it will immediately send the mobile phone newsletter to σ σ - (4 0 ) or medical staff (5 0). On the multi-image processing server (2 η ) μ ΑΑ V d U ), as shown in the figure. We arranged the multi-grid face above the gold J, and serialized it in 200919382 as the nickname camera, Μ camera. At the bottom of the screen (4), there are 9 kinds of "like falls, the type of the pattern for identification. The degree of similarity after the comparison of the face and the fall model is listed next to each image. The smaller the number, the more similar. When continuous When the pictured face is similar to the same fall type, the system will issue a mobile phone warning message 'The picture is shown in the third picture. On the eve, the next time, the fall identification algorithm of the single image of the system and the multi-image string Flow Acceleration Algorithm. A 'fall image recognition algorithm: The fall shirt image recognition is divided into four steps, which are: Step 1: Image capture thread
系統啟動前，先留存_萨I 存張热人場景圖（6 0 )，如第 A圖所不，供之後去 拍攝現場晝面（7 二田^先啟動後’以固定速率 s . 如弟四B圖所示，並將該現場晝 面儲存至該記憶體結構内； 步驟—：影像處理執行緒 此步驟的主要工作，3從τ目p去 ^ Λ ^ ^ e 疋將現场晝面（70)中與原本 :〜圖（60)之間差異的 處理的動作過濾出來，且確伴物…8 〇 )利“像 所示）。 隹保物體不會失真（如第五A圖 步驟三：人形輪摩萃取執行緒 此步驟的工作，县收旦/ 丄 〇)描纷出來，做為Γ (80)的輪廊（9 份人形輪廊（90)越：：斷（如弟五B圖所示）。這部 越元整不斷裂，判斷的結果越準確。 200919382 步驟四：跌倒樣型識別執行緒 主要將現場晝面（7〇)的人形輪 類似跌倒的樣型作比對 υ ) y張 對每9張類似跌倒的 安、跪姿、與坐姿三大類。而這些樣型都 = 同的處理步驟取出輪廊特徵點並儲存於系統中^以= 晝面（7〇)的人形輪^9Q)與這些特徵點逐一比2 即可根據比對結果判讀跌倒與否。 、 上述四個步驟會循環進行，如 訊息，且累計次數超過—閥值，代面都出現跌倒 代表旦面中的人員發生胜 :外的機率很高’因此發出手機簡訊通知醫護 0 )。整個程序可以第六圖所示。 、 一、夕格影像串流加速計算演算法： 當系統外接多台攝影機（10)在不同場景進行跌倒 制時’每一台攝影機（10)都會產生影像申流，而每 一運串流都會經過上述四個運算步驟：影㈣取⑴峨 Fetch)執行緒、影像處理（image ^咖叫）執行緒、人 形輪廓萃取（Sequence GeneratlQn)執行緒、跌倒樣型識別 (PaUen Re⑶gnitlQn)執行緒。丨了後面教述方便，我們 將這四步驟依序簡稱I F、I P、S G、與p R。 為了加速多格影像串流的運算，我們採用多執行緒 (Multi threading)與管線平行運算（pipelining)的原理， 由每—執行緒負責單-步驟的工作’而所有執行緒則依據 :線化的排程來運作，如第七圖所示。當第一台攝影機的 第一串流影像資料進入系統後，該影像擷取執行緒（丨F ) 200919382 :執行影像掏取的運算，待運算完成後，擷取的影像會被 玄記憶體結構中特定的資料結構内。接著由該影像處 =行緒"”接手進行影像處理的工作。在此同時， :“象砧取執行緒（1f)立即接收第二台攝影機的第二 =像串”料，重複影像擷取的工作。等該影像#貞取執行 二（I F )與該影像處理執行緒（I P )平行運算完畢後， f衫像串流資料輪由該人形輪靡萃取執行緒（§ G )運 鼻’該第：影像串流資料交由該影像處理執行、緒（I p ) 第—影像串流資料則由另一新接收的第-影像操 仃、· 1 F )運算，所以在這個時間點時，系統等於 :個執行緒平行進行運算。以此類推，系統中最高可同 古予在四個平仃運算的執行緒。如此排程的結果，系統合 ^穩定的產能（ThrQughput)來處理^攝影機隸 串流’達到多工處理以加速影像運算之目的。 〜像 ^明所涵蓋的跌倒辨識演算法與多格影像串流加速 异肩异法可輕易實作於個人電腦的標準架構 獲得良好的效能表現。而且利用前述之演算法不僅可應用 於銀t族跌倒照護且不影燮+ 心 為智慧型居家保全、㈣：：年广曰…更可以開發 偵測.、物入侵等相關的產業應用。 總結而言，本系統利用影像處理與樣型識別技術完成 跌倒偵測照護之服務’崎無須佩帶任何微感測器於身 。其次’本發明利用多執行緒與管線化排程的原理，完 =機四t影像處理之平行運算工作。再者，本發明在辨 硪出可能發生人員跌倒畜杜吐 、失丫 i事件％，即時傳送警告簡訊給警衛 14 200919382 或醫護人員前往察看，使得照 # ^ m μ κ ^ 有政率。此外，本發明 係如用個人化電腦標準架 热須另外附加特定之軟 -,^ y 知月所開發之跌倒影像辨識演算 利用多台攝影機拍攝不同 靡特徵點並炉類跌倒，，之媒’旦面，即時操取人形輪 理… 糊之樣型進行比對，可大幅簡化影像處 且不會影響辨識正確率或誤判率。而且本系統 影像串流加速計算演算法，可大幅改 像處理效能等多種功效。 ::係針對本發明之可行實施例為本發明之技術特徵 惟’熟悉此項技術之人士當可在不脫離本 更=精神與原則下對本發日月進行變更與修改，而該等變 二改」皆應涵蓋於如下申請專利範圍所界定之範疇中。 【圖式簡單說明】 圖：係本發明可行實施例之實施架構示意圖。 第:圖.係本發明可行實施例之實際操作晝面示意圖。 第一圖，係本發明實施例之傳送手機警告簡訊示意圖。Before the system starts, first save the _ Sa I save the hot scene map (60), as shown in Figure A, for later to shoot the scene after the surface (7 Ertian ^ first start after 'at a fixed rate s. As shown in Figure 4B, the surface of the scene is stored in the memory structure; Step-: Image Processing Execution The main work of this step, 3 from τ目p to ^ Λ ^ ^ e 疋 will be the scene (70) The action of the difference between the original and the original: ~ (60) is filtered out, and it is indeed accompanied by ... 8 〇) profit "as shown." Ensure that the object is not distorted (as in step 5A) Three: The work of the humanoid wheel and the extraction of this step, the county collection / 丄〇) is drawn out, as the porch of the Γ (80) (9 copies of the humanoid porch (90):: broken (such as the young five Figure B). The more the yuan is not broken, the more accurate the judgment is. 200919382 Step 4: The fall type recognition thread mainly compares the humanoid wheel of the scene (7〇) with a similar fall type. υ ) y Zhang is for every 9 similar falls, 跪 position, and sitting posture. These types are the same processing steps to take out the features of the porch. Point and store in the system ^ ^ 昼 (7〇) humanoid wheel ^9Q) and these feature points one by one 2 can be judged according to the comparison results fall or not., the above four steps will be repeated, such as messages And the cumulative number of times exceeds the - threshold, the generation of the face appears to fall on behalf of the person in the face of the victory: the probability of outside is high 'so the mobile phone newsletter to inform the medical care 0. The entire program can be shown in the sixth figure. Xiage video stream acceleration calculation algorithm: When the system externally connects multiple cameras (10) in different scenes, the falling system of each camera (10) will generate image flow, and each stream will pass the above four Operation steps: shadow (4) take (1) 峨 Fetch) thread, image processing (image ^ coffee) thread, sequence contour extraction (Sequence GeneratlQn) thread, fall type recognition (PaUen Re (3) gnitlQn) thread. For convenience, we will refer to these four steps as IF, IP, SG, and p R. To speed up the operation of multi-image stream, we use multi-threading and pipeline parallel operation (pipelin The principle of ing) is that each thread is responsible for the single-step work' and all threads are operated according to the linearized schedule, as shown in Figure 7. When the first camera is the first stream image After the data enters the system, the image capture thread (丨F) 200919382: performs an image capture operation, and after the operation is completed, the captured image is subjected to a specific data structure in the meta memory structure.处=行绪"" took over the work of image processing. At the same time, "the anvil picking thread (1f) immediately receives the second camera image of the second camera, repeating the image capture work. After the image #贞Execution 2 (IF) is paralleled with the image processing thread (IP), the f-shirt stream data wheel is extracted by the humanoid rim (§ G). The video stream data is transferred to the image processing execution, and the first (Ip) video stream data is calculated by another newly received first image operation, · 1 F ), so at this time point, the system is equal to : A thread executes the operation in parallel. By analogy, the highest achievable thread in the system can be used in the four flat operations. As a result of this scheduling, the system combines the stable capacity (ThrQughput) to handle the camera stream to achieve multiplex processing to accelerate image calculation. ~ The fall recognition algorithm and multi-image stream acceleration as described by Ming Ming can be easily implemented in the standard architecture of personal computers to achieve good performance. Moreover, the above-mentioned algorithm can be applied not only to the silver t-family fall care but also to the heart-warming home security, (4):: the year-wide 曰 ... can also develop related industrial applications such as detection and intrusion. In summary, the system uses image processing and pattern recognition technology to complete the service of fall detection and care. Saki does not need to wear any micro-sensors. Secondly, the present invention utilizes the principle of multi-threading and pipelined scheduling, and completes the parallel operation of the four-t image processing. Furthermore, the present invention identifies a possible occurrence of a person falling down, a missed event, and immediately sends a warning message to the guard 14 200919382 or the medical staff to view it so that the # ^ m μ κ ^ has a political rate. In addition, the present invention is to use a personal computer standard rack heat to be additionally attached to a specific soft-, ^ y zhi zhiyue developed fall image recognition calculation using multiple cameras to shoot different 靡 feature points and furnace falls, the media ' Once the face is taken, the humanoid wheel can be manipulated in real time... The comparison of the paste type can greatly simplify the image without affecting the recognition accuracy or false positive rate. Moreover, the system video stream acceleration calculation algorithm can greatly improve various effects such as processing performance. The following is a description of the possible embodiments of the present invention. However, those skilled in the art can change and modify the present day and the month without departing from the spirit and principle. The changes should be covered in the scope defined by the scope of the patent application below. BRIEF DESCRIPTION OF THE DRAWINGS Fig.: A schematic diagram of an implementation architecture of a possible embodiment of the present invention. Fig. is a schematic view showing the actual operation of a possible embodiment of the present invention. The first figure is a schematic diagram of a transmission mobile phone warning message according to an embodiment of the present invention.
势 A 四圖·係本發明可行實施例之無人背景示意圖。Potential A Four Diagrams is a background diagram of an unmanned background of a possible embodiment of the present invention.
第 P — 圖：係本發明實施例之現場實際拍攝示意圖。 =五A圖：係本發明實施例之去背景後之人形示意圖。 第五B圖.係本發明實施例之人形輪靡特徵示意圖。 第六圖：係本發明實施例之跌倒偵測辨識流程示意圖。 第七圖：係本發明可行實施例之多執行緒與管線平行 運算示意圖。 【主要元件符號說明】 15 200919382 (1 0 )攝影機 （2 (3 0 )保全端警示裝置 （4 (5 0 )醫護人員 （6 (70)現場畫面 （8 (9 0 )輪廓 (I F )影像擷取執行緒 (I P )影像處理執行緒 (S G )人形輪廓萃取執行緒 (P R )跌倒樣型識別執行緒 )多格影像處理伺服器 )警衛 )無人場景圖 )物體 16P - Figure: is a schematic diagram of the actual shooting of the scene in the embodiment of the present invention. = Figure 5A is a schematic diagram of the human form after the background of the embodiment of the present invention. Fig. 5B is a schematic view showing the characteristics of the humanoid rim according to the embodiment of the present invention. Figure 6 is a schematic diagram of the fall detection identification process in the embodiment of the present invention. Figure 7 is a schematic diagram of parallel operation of multiple threads and pipelines in a possible embodiment of the present invention. [Main component symbol description] 15 200919382 (1 0 ) Camera (2 (3 0) security alert device (4 (5 0 ) medical staff (6 (70) live screen (8 (9 0) contour (IF) image撷Take thread (IP) image processing thread (SG) humanoid contour extraction thread (PR) fall type recognition thread) multi-image processing server) guard) unmanned scene map) object 16