1307061 九、發明說明： 【發明所屬之技術領域】 本發明係關於一種用於偵測事件之影像監控方法及系 統，尤其關於一種可經客製化以偵測使用者所想要偵測的 事件之即時影像監控方法及系統。 【先前技術】 因相關數位系統技術的進步，影像監控系統由以往之類1307061 IX. Description of the Invention: [Technical Field] The present invention relates to an image monitoring method and system for detecting an event, and more particularly to an event that can be customized to detect a user's desired detection Instant image monitoring method and system. [Prior Art] Due to the advancement of related digital system technology, the image monitoring system has been used in the past.
比監視系統，逐漸為數位監視系統所取代，而其中又分為 單機型架構及以個人電腦錢器為基礎之架構兩大數:監 視系統類型。 單機型架構數位監㈣統以單晶片、特定功能冗或· 搭配韌體及其他儲存設備運作’具有簡單、穩定、功能單 -等特色，因而較適合單調的、長時間性、穩定性高的環 境監看與錄影存料應用場合，然而其缺乏對監看環境之 自動化分析能力’因此若有全時間預警之需求時仍須配置 人力配置協助監看。 富彈性與實用性 而以個人電腦伺服器為基礎之數位監視系統，因有庫用 程式u於開發性與通用替換性，加以網路功能及介面擴 充功此強大，以及具有基礎的影像内容分析能力，因此較 然而 . j白知之以個人電腦伺服器為基礎之數位監視 糸統，其使用模式雖可透過程式設計分析監控畫面，卻需 要系統開發廉商針對所欲警示情景之畫面情況預先分析， 並將所分析之結果特徵預載於系統資料庫中，以供日後與 115138.doc 1307061 所揭取之監視晝面比對使用；當比對相符時，即可針對所 欲警示情景自動發出通報。然而，這種仰賴廠商預先研發 並輸入比對條件之使用模式，不僅無法適用於各種變化快 速、複雜多樣化的環境變動情形，若客戶有客製化的警示 情景债測需求時，也必需再次委託廠商針對需求重新開發 而無法即時施行。此使用模式不僅缺乏應用彈性，同時也 增加系統開發成本。舉例來說，中華民國專利公告第 1235964號發明專利*絲姐一 揭不—種偵測一預定區域中之早 期火f生成之方法及影像監視系統。然而此先前技術受限 於僅能偵測特定之火災拿杜 事件，其偵測特定之火災事件的判 断條件係固定而無法針對使用實地環境作客製化㈣整。 =’仍需求一種影像監控系統可彈性地用 用者想要偵測的事件。 』仗 【發明内容】 ^ 貞測事件之影像監控方法及系统。 於可針對使用者所想要情測的事件預先建立可代 表》亥事件之特徵向量時變模， 若經比對所監控之影像的特徵/即時監控影像時’ 可代表該事件之特徵向旦日士#/夏時間函數與該所預建之 事件之發生。 里^模型相似時，即可偵測出該 件監控方法及系統，其事件之發生的判斷停 件係基於針對❹者射m ，、 事件之特徵向量時變^事件預先建立可代表該 單、變化紐人，而建立該時變模型的方式簡 °夕、富選擇性與彈性，故人人皆可自行設定 II5I38.doc I3〇7〇6] 或隨時更改’具時效性’且客製化應用程度極高。且可達 成真正無人化之自動監控目#，或作為現行監控系統之辅 助機制’大幅減少人力支出成本。 根據本發明之-態樣’-用於_事件之影像監控方法 包含下列步驟：取得影像步驟，其取得在連續複數個不同 時間點所拍攝的複數張影像；擷取特徵向量步驟，，取 出該等影像中每一影像的至少一特徵向量，以獲得該等影 像之該至少一特徵向量的一時間函數；及判斷步驟，其藉 由比對該至少-特徵向量的該時間函數及對應—事件之^ 至少-特徵向量的一時變模型’判斷該事件是否已在該； 時間點内發生。其中，對應該事件之該至少—特徵向量的 該時變模型係預先經由-預建模型步驟而獲得，該預建^ 型步驟包含下列步驟：-使用者針對該事件選擇要拇取的 該至少-特徵向量；該使用者提供至少—組複數張預選影 像’其中每-組複數張預選影像中均紀錄有該事件之發 生；分別從該至少-組複數張預選影像的每一影像中擷： 出該至少一特徵向量，以獲得至少一組該至少一特徵向量 的時間函m析該至少-組該至少—特徵向量的時= 函數’以產生可代表該事件的該至少—特徵向量的該時變 模型。 根據本發明之另一態樣，一用於偵測事件之影像監控系 統包含：-影像攝取裝置，其在連續複數個不同時間點攝 取複數張影像；及一影像分析裝置，其耦合至該影像攝取 裝置以取得該等影像。其中’該影像分析裝置經組態以： 115138.doc 1307061 擷取出該等影像中每一影像的至少一特 箄导彡榇* μ 以獲得該 寺〜像之碟至少一特徵向量的一時 至小_ 4dt 夂稽由比對該 夕—特徵向量的該時間函數及對應一 料a旦 π n s哀至少一特 二里=時變模型’判斷該事件是否已在該等時間點内 建才1^、中，該影像分析裝置係預先經組態以執行一預 建才吴组動作而產生對 變模型，h 件該至少一特徵向量的該時 亩灿^ 勒卞取侍使用者針對該 入 主^特徵向篁’讀取使用者所輸 入至少—組複數張預選影像，其中每一 中妁+ 八Y母，·且複數張預選影像 景Μ象的备 ，刀别從該至〉、一組複數張預選 ❼像的母一影像中擷取 一 特徵向罝，以獲得至少 :該至少-特徵向量的時間函數；及分析該至少一組該 至乂一特徵向量的時間函數，以 人 少一牿η Α θ t 屋生了代表該事件的該至 ^ 将Μ向ϊ的該時變模型。 【實施方式】 以下將利用所附之圖式針對本發明之且體膏祐γ 細陳述之，其中並干出㈣士 之』實施例加以詳 士 出根據本發明之較佳具體實施例。鈇 而應S亥注意的是，太路 …、 11刼用許多不同之形式來實施， 並不又限於此處陳述之較佳具體實施例。 根據本發明之影俊ε 時監於fM象日# ±息1 工方法及系統之具體實施例，在即 了監控影像時，主要係 用者想要偵測的事件是…先建之_而判斷-使 由下列步驟獲得： 精 首先’利用圖像識別技術 何針對一或數組紀錄有一特定事 115138.doc 1307061 件之發生過程的連續數張影像作 像中的至少一個特徵向量’掏取出個別單張影 續數張影像的-或數組特徵向量二=-或數組連 量可包含例如：某特定物件大小/ ==°此特徵向 木方位、紋理結構等常見的物件描 色月 描述方式，亦可為影像的平 二一特殊的 等。影像的各種特徵向量及其料方J；度、尚頻成份比例 像處理/電腦視覺專書（Λ〇Μ’可藉由在習知影 processing·，ISBN:〇 % Dlgltal 酿ge v. · , Τΐ 8 Shapir^ '2001. CompUterThe ratio monitoring system is gradually replaced by a digital monitoring system, which is divided into a single-machine architecture and a structure based on a personal computer money device: the type of monitoring system. The single-machine architecture digital monitor (4) is characterized by single-chip, specific function redundancy, or with firmware and other storage devices. It has simple, stable, single-function and so on, so it is more suitable for monotonous, long-term, high stability. Environmental monitoring and video storage applications, however, lack the ability to automate the analysis of the monitoring environment. Therefore, if there is a need for full-time warning, human configuration must be configured to assist in monitoring. Flexible and practical, digital computer-based digital surveillance system, with the development and universal replacement of library applications, network functions and interface expansion, and basic image content analysis The ability, therefore, is more than a personal computer server-based digital surveillance system. Although its usage mode can analyze the monitoring image through programming, it requires the system to develop a low-cost pre-analysis of the image of the desired warning scenario. , and the characteristics of the analyzed results are pre-loaded in the system database for later comparison with the surveillance surface exposed by 115138.doc 1307061; when the comparison is matched, the alarm situation can be automatically issued for the desired warning situation. Notification. However, this kind of usage pattern that relies on the manufacturer's pre-development and input of comparison conditions can not only be applied to various environmental changes with rapid, complex and varied changes. If customers have customized warning scenarios, they must also be re-evaluated. The commissioned vendor is redeveloped for demand and cannot be implemented immediately. This usage model not only lacks application flexibility, but also increases system development costs. For example, the invention patent of No. 1235964 of the Republic of China Patent No. 1235964 discloses a method for detecting the early fire generation in a predetermined area and an image monitoring system. However, this prior art is limited to detecting only specific fires, and the conditions for detecting specific fire events are fixed and cannot be customized for use in the field environment (4). =' There is still a need for an image surveillance system that can flexibly use the events that the user wants to detect. 』仗 [Summary] ^ Image monitoring method and system for speculative events. The model can be pre-established for the event that the user wants to test, and the feature vector representing the event can be changed. If the image of the monitored image is compared/simultaneously monitored, the feature can represent the event. The Japanese/# summer time function and the pre-built event occurred. When the model is similar, the monitoring method and system can be detected. The judgment of the occurrence of the event is based on the shot of the target, and the event vector of the event is pre-established to represent the order. Change the newcomers, and the way to establish the time-varying model is simple, selective and flexible, so everyone can set II5I38.doc I3〇7〇6] or change the 'time-sensitive' and customized application at any time. Extremely high. And it can be turned into a truly unmanned automatic monitoring target #, or as an auxiliary mechanism of the current monitoring system', which greatly reduces the cost of labor. The image monitoring method for the event according to the present invention includes the following steps: obtaining an image step of acquiring a plurality of images taken at successive plurality of different time points; taking a feature vector step, and extracting the image And at least one feature vector of each image in the image to obtain a time function of the at least one feature vector of the images; and a determining step of comparing the time function of the at least-feature vector with the corresponding event ^ At least - a time-varying model of the feature vector 'determines whether the event is already there; occurs within the time point. Wherein the time-varying model corresponding to the at least-feature vector of the event is obtained in advance via a pre-built model step, the pre-built step comprising the steps of: selecting, by the user for the event, the at least the thumb a feature vector; the user provides at least a plurality of sets of preselected images in which each of the plurality of preselected images records the occurrence of the event; respectively, from each of the at least one set of preselected images: Deriving the at least one feature vector to obtain a time function of at least one set of the at least one feature vector, extracting the at least one set of at least one time = function ' of the feature vector to generate the at least one feature vector representative of the event Time-varying model. According to another aspect of the present invention, an image monitoring system for detecting an event includes: an image capturing device that takes a plurality of images at successive plurality of different time points; and an image analyzing device coupled to the image The device is ingested to obtain the images. Wherein the image analysis device is configured to: 115138.doc 1307061 撷 extract at least one special guide 彡榇* μ of each image in the images to obtain at least one feature vector of the temple-image disk _ 4dt 由 由 对该 对该 — — — — — — — — — 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征 特征The image analysis device is configured in advance to perform a pre-construction action to generate a pair of models, and the at least one feature vector of the at least one feature vector is for the user. The feature 篁 'reads at least one set of pre-selected images input by the user, wherein each middle 妁 + eight Y mother, and a plurality of pre-selected image scenes are prepared, the knives are from the ～, a group of plural Extracting a feature direction from the mother image of the preselected key image to obtain at least: a time function of the at least-feature vector; and analyzing a time function of the at least one set of the one feature vector to be less than one person η Α θ t The house gave birth to the representative of the event The Μ ^ to the time varying model of ϊ. [Embodiment] Hereinafter, the embodiment of the present invention will be described in detail with reference to the accompanying drawings, in which the embodiments of the invention are described in detail.鈇 S 亥 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 注意 、 、 、 、 、 、 、 、 、 、 、 。 。 。 According to the specific embodiment of the method and system of the fM image day, in the case of monitoring images, the main user wants to detect the event is... - Obtained by the following steps: Fine first 'Using image recognition technology for one or array of records has a specific event 115138.doc 1307061 piece of the process of the process of at least one feature vector of the image of the process' Zhang Ying continued several images - or array feature vector two = - or array linkages can include, for example: a specific object size / == ° This feature is described in the wood orientation, texture structure, and other common objects, or The image of the flat two special and so on. The various feature vectors of the image and its material side J; degree, frequency component ratio like processing / computer vision book (Λ〇Μ' can be processed by 影影, ISBN: 〇% Dlgltal ge ge v. · , Τΐ 8 Shapir^ '2001. CompUter
Vision . Upper Saddle River, New Jersey P ,· 如.)或學術期刊文獻中（如P_rn ^ ―㈣， 現，亦可由使用者或廢商自行開==技術來實 中。 π间贫加入並儲存於資料庫 处分析該-或數組特徵向量對時間的函數以得到最 + 夕特徵向罝的—時變模型（例如， 精由統計方法找出該至少一特徵向量隨時間改變的趨勢或 規則）。該特徵向量的時變模型可以例如多項式逼近 (P〇lyn〇imaUurve mting)函數或其它方式紀錄下來。 此處代表-事件的特徵向量時變模型，可為廠商在生產 即時監控系統時，預先針對特定事件及其常遇到的環境變 化藉由上述步驟分析後所建立且預先儲存至系統中，供使 用者決定是否啟動其中之預設模式選項進行監控分析。例 如，廠商可針對火災發生時的場景預先分析，將此特徵輸 115I38.doc 10 1307061 只竹琿〒，供使用者 況。 另方面’因應使用者在不同璟掊,(·主、σ > 下的特殊使用場人祛田土 义月况或各種自訂條件 擇要心 者亦可預先針對所欲谓測事件選 擇要#緣的特徵向量。針對使用者所輪入-或數組 測事件之發生過程的連續數張影像（其電 =二建立能代表使用者所欲_事件的特徵向量 變=其中如何針對所則貞測事件來選擇要擷取的特 ^自=經由專家教育訓練、使用者自行嘗試驗 -(自爾模擬影片測拭)，或使用者經驗交 必^主意的是，隨著選擇要擷取的特徵向量越多'或特徵 ^篁對該情況的描述越貼切，則系統的準確度即相對提 Ν、㈣率降低’料算複誠㈣對提高。 在藉由以上方式預先建立之能代表使用者所欲谓測事 :特徵向量的時變模型後，在即時監控影像時可針對即時 攝取的連續數張影像，擷取出所選取的特徵向量以即 時的特徵向量對時間的函數，並將該時間函數與該預切 ^之此代表使用者所則貞測事件的特徵向量的時變模型比 較’即可判斷出該事件是否發生；或者計算出該事件發生 的機率’而由使用者自行設定當其判斷出的機率高於” 百綱則判斷該事件發生。此處所用之比較判斷的方 式’可採用έ吾音訊自卜卜抖社 飞心比對技術（例如，動態比對法 糊邏輯（fuzzy l〇glc)或其它習知技術。 、、 115138.doc 1307061 ♦田判斷出該事件已發生時，使用者可依需要自動執行適 、:的動作’例如：產生一聲音、燈光或電性警示信號，或 :過企業内部網路（i咖nei)或網際網路（i咖⑽)發出警告 電子郵件給使用者或管理者、透過無線網路發出警告短訊 (SMS)或警告電話呼叫（ph〇ne Q⑴給 甚至積極地主動阻止該事件之繼續發生(例如 事件的么生可自動灑水），藉此達成全自動監控的功能。 ，理上述根據本發明之—影像監控方法，其具體實 包含如下步驟： ⑴：用者預先針對所欲谓測事件(例如火災，或人為的特 ^打為’如徘彻或攀爬建物）選擇要摘取的至少一個特 徵向量； ' (2)使用者預先提供—或數組紀錄有該事件發生過程的複 數張連續影像，該等連續影像可為以攝影機攝取該事 件的真實發生過程或以電腦模擬該事件； ⑺預先藉由影像處理技術，逐一對該等連續影像中的每 :帖：像進:亍處理(例如，色彩空間轉換、灰階或二值 呆 雜5孔濾除、線段擷取 '區域分割、八 =徵描述法)，藉以擷取出各影像的特徵SI: =:之類型及靜態狀態。接著，將所有影像中的 ’依時間順序排列，分析特徵向量隨時間的 -化趨勢(代表狀態的變化)’並依此得 =特徵向量時變模型，作為即時監 = 依據，並紀錄下此特徵向量時變模型。(舉例來π 115138.doc 1307061 火災發生時’火苗有擴展變大 勢，因此，可以火苗的頂點位置，八本売度增加的趨 均亮度作為特徵向量，當隨時間增：像平 置升高，分布面積變大，且平均2時’即有頂點位 其特徵向量的時變模型。）：A s加的趨勢作為 (4)於即時監控影像系統運作時，取 時間點所拍攝的複數張續複數個不同 影像的特徵向量，以獲等影像中每一 αΛ ^ 獲于該等影像之特徵向量對暗門Vision. Upper Saddle River, New Jersey P, ·.. or in academic journal literature (such as P_rn ^ ― (4), now, can also be opened by the user or the waste business == technology to implement. π lean poor to join and store The function of the - or array feature vector versus time is analyzed at the database to obtain a time-varying model of the most + eigen-wise ( (for example, the statistical method is used to find the trend or rule of the at least one eigenvector changing with time) The time-varying model of the eigenvector can be recorded, for example, by a polynomial approximation (P〇lyn〇imaUurve mting) function or other means. Here, the eigenvector time-varying model representing the event can be pre-arranged by the manufacturer in the production of the real-time monitoring system. The specific events and their frequently encountered environmental changes are established by the above steps and stored in the system in advance, and the user can decide whether to activate the preset mode option for monitoring analysis. For example, the manufacturer can target the fire. The scene is pre-analyzed and this feature is exported to 115I38.doc 10 1307061 for the user's condition. In addition, 'in response to the user's different experience, (· σ > Under the special use of the field, the people of the field, or the various custom conditions, you can also select the feature vector of the #缘 for the desired event. For the user to round - or array Measure several consecutive images of the occurrence of the event (the electric=two establishes the eigenvectors of the event that the user wants to _the event = how to select the specific traits to be taken for the estimator event = through expert education training The user will try it on his own - (self-simulating film test), or the user experience must be, the more the feature vector is selected, or the more appropriate, the more appropriate the description of the situation is. Then, the accuracy of the system is relatively high, and (4) the rate is reduced. [Recommended by Fucheng (4). In the above-mentioned way, it can represent the user's desired measurement: the time-varying model of the feature vector, after When the image is monitored in real time, the selected feature vector can be extracted as a function of the instantaneous feature vector versus time for the consecutive images captured in real time, and the time function and the pre-cutting event represent the user's speculative event. of The time-varying model of the feature vector is compared with 'can determine whether the event occurs; or the probability of occurrence of the event is calculated' and is set by the user to determine that the probability is higher than the one hundredth. The method of comparative judgment used by the premises can be used by the έ 音 自 卜 ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ( ♦ When the field judges that the event has occurred, the user can automatically perform the appropriate action as needed: for example, generate a sound, light or electric warning signal, or: through the internal network of the enterprise (i cafe nei) or the Internet The Internet (Ica (10)) issues a warning email to the user or administrator, sends a warning message (SMS) or a warning phone call over the wireless network (ph〇ne Q(1) to even actively proactively prevent the event from continuing ( For example, the event can be automatically sprinkled, thereby achieving the function of fully automatic monitoring. The above-mentioned image monitoring method according to the present invention specifically includes the following steps: (1): The user pre-measures the event to be described (for example, a fire, or an artificial special action such as smashing or climbing a building) Selecting at least one feature vector to be extracted; ' (2) the user provides in advance - or an array records a plurality of consecutive images of the event occurrence process, and the continuous images may be the actual occurrence process of taking the event by the camera or The computer simulates the event; (7) by image processing technology, one by one of each of the consecutive images:: image: 色彩 processing (for example, color space conversion, grayscale or binary value 5 hole filtering, line segment Take the 'area segmentation, eight = sign description method), so as to extract the characteristics of each image SI: =: type and static state. Next, the 'in chronological order of all images, analyze the trend of the feature vector over time (representing the change of state)' and get the eigenvector time-varying model as the basis of the real-time monitoring, and record this Eigenvector time-varying model. (For example, π 115138.doc 1307061 When the fire occurs, the flame has a tendency to expand and become larger. Therefore, the apex position of the flame, the average brightness of the increase of the eight twists can be used as the feature vector, and when it increases with time: the image rises, The distribution area becomes larger, and the average time is 2', that is, the time-varying model with the eigenvectors of the vertex bits.): The trend of A s is added as (4) when the real-time monitoring image system operates, taking the multiple shots taken at the time point a plurality of eigenvectors of different images to obtain each Λ ^ in the equal image obtained from the eigenvectors of the images
令函數；及藉由比對該特旦 S 先建立的對應該事件之特如^該時間函數及該預 大於一…主 時變模型，若相似度 生。 】斲該事件在該等時間點内發 此⑷展示根據本發明之-較佳具體實施例。在Let the function; and if the similarity is generated by comparing the time event of the corresponding event to the special time, such as the time function and the pre-more than one... main time-varying model. The event is issued at these points in time (4) showing a preferred embodiment in accordance with the present invention. in
Γ:::、者欲_的事件為··有人走進監控區 傻轉下起身、並離開龄扯P 欲擷取的特徵向量為移動：：用者針對此事件選取 「骨的「骨架長短軸比值」及 件；軸夾角值」，並提供-組事先錄製的包含此事 ==複數張影_⑷至I(h))。接著 :移動：體，並進行適當的影像前置處理(例如··去雜 r Λ ^ 戈連、，·。等），以便將該移動物體的輪 廓找出麵-步地，取出各影像中移動物體之骨架資訊 il* =至1(13)所不）並進而榻取出各影像中該移動物體的 骨条之對應長短輛之夹角與長度比等特徵向量。最後紀錄 li5138.doc -13- 1307061 了士圖1 (q)中所不特徵向 itb a： ^ ^ ^ 釕盼間的函數，曲線11即為 此事件發生時移動物體的「 巧 . 朱長紐軸比值J對時間的Μ 係，而曲線12即為此事件發 了頂的關 夾肖侑Ρ 件發生4移動物體的「骨架長短軸 文角值」對知間的關係、’其可分別用如. (polynomial curve fitti ) 夕式迴近 中，為简㈣四 §)函數加以紀錄儲存。在此實施例Γ:::, the event of the person wants to be _· Someone walks into the surveillance area and turns to get up, and leaves the age to pull P. The feature vector to be captured is:: The user selects the skeleton length of the bone for this event. Axis ratio "and piece; axis angle value", and provides - group pre-recorded contains this thing == plural picture _ (4) to I (h)). Then: move: body, and perform appropriate image pre-processing (for example, ······························································· The skeleton information of the moving object il* = to 1 (13) does not) and then the feature vector such as the angle and the length ratio of the corresponding long and short vehicles of the moving object in each image is taken out. The last record li5138.doc -13- 1307061 is a function of the characteristic of itb a: ^ ^ ^ in the figure 1 (q), curve 11 is the object that moves the object when this event occurs. The axis ratio J is related to time, and the curve 12 is the top of the event. The relationship between the "skeletal length and the short axis value" of the moving object is 4, and the difference can be used separately. For example, in the polynomial curve fitti, the simple (four) four §) functions are recorded and stored. In this embodiment
'' 說月，僅輪入一組包含此畜杜| 4、_R 浮罢彡择 ^ 3此事件發生過程的複數 張景/像，並以此分析得到的特 * 0里對犄間的函數作為代 表此事件之時變模組。 〜 當曰後需進行監控此項行為. 訊，同樣選取「骨架長短轴二Π:對輸入之視 G Μ J及「骨架長短軸夾角 值」作為特徵，藉由如習知的立 @1. ,, 〇曰讯息比對技術一動態比 對法則（dynamlc programmi )，告 ώ 丄 g)田輸入之連續影像特徵曲 線與内存的特徵模型曲線相似 1度違一定比例以上時（例如 7〇%，可由使用者自行訂定）， )系統即判斷為疑似警示之行 為Γ現，進而回應相對輸出（如警報或郵件），達成預警 功月b。在此實施例中，亦可再知μ教去， J再加上移動物體的「骨架長軸 線與水平線的夾角」作為另—個牯 1U特徵向量，即能進一步細 判S亥移動物體係處於站立或倒臥之狀熊。 圖2⑷至2(δ)展示根據本發明之另—較佳具體實 在此實施例中，使用者欲偵測的事件為：監控區域範圍中 發生了火災。因為火災通常同時伴有 了旰’煙霧及火焰等現象的 出現’針對此情況，使用者選取欲梅取的特徵向量為監押 區域範圍中變化物體(相對於各時間點均相同之不變背旦） 之「平均色度值」及「高頻能量相對於全頻能量的: 115138.doc -14. 1307061 值」，此選擇係基於火焰發生時之平均色度值有其特定範 圍（可為橙、紅或藍）’因此可用做標定火焰；而煙霧產生 時，其顆粒濃度阻礙光線穿透，因此阻擋對背景的觀察， 此現象將反映在影像中高頻成分的降低（能量向低頻移 動）。因此由觀察高頻能量相對於全頻能量的比值之特徵 向量，將可有效標定煙霧。 使用者提供一組事先針對該監控區域所錄製的包含'' Say the month, only round into a group containing this animal Du | 4, _R floating choice ^ 3 the occurrence of the process of the multiple scene / image, and the analysis of the special * 0 in the inter-day function As a time-varying module that represents this event. ~ When you need to monitor this behavior, you can also select the "frame length and length axis two: the input view G Μ J and the "frame length and length axis angle value" as characteristics, by the conventional standing @1. ,, 〇曰 message comparison technique-dynamlc programmi, caution ώg) The input image characteristic curve of the field input is similar to the characteristic model curve of the memory when the degree is more than a certain percentage (for example, 7〇%, It can be set by the user.) The system judges the behavior of the suspected warning, and then responds to the relative output (such as alarm or mail) to reach the warning power b. In this embodiment, it is also possible to know that μ is taught, and J is added to the "angle between the long axis of the skeleton and the horizontal line" of the moving object as another 牯1U feature vector, which can further judge that the mobile system is in A bear standing or lying down. Figures 2(4) through 2([delta]) show another embodiment in accordance with the present invention. Preferably, in this embodiment, the event to be detected by the user is: a fire has occurred in the area of the monitored area. Because the fire is usually accompanied by the appearance of 烟雾 'smoke and flames'. In this case, the user selects the feature vector taken by the plucking to be the changing object in the scope of the custody area (the same invariable time relative to each time point) ) "average chromaticity value" and "high frequency energy relative to full frequency energy: 115138.doc -14. 1307061 value", this selection is based on the average chromaticity value of the flame when it has a specific range (can be orange) , red or blue)' can therefore be used to calibrate the flame; when the smoke is generated, its particle concentration hinders the light from penetrating, thus blocking the observation of the background, which will reflect the decrease in the high frequency component (the energy moves to the low frequency) in the image. Therefore, by observing the characteristic vector of the ratio of the high frequency energy to the full frequency energy, the smoke can be effectively calibrated. The user provides a set of records recorded in advance for the monitored area
件發生過程的複數張影像圖2⑷至2(χρ在此實施例中， 該等像係真實攝錄使用者自行模擬的示範性火災之結果， 在該示範性火災發生期間，曾受風吹襲影響，丨中圖2(g) 及2(h)中紀錄有明顯火苗起始點，圖2(〇中紀錄因風吹 焰受影響的影像。 接者’針對所輪入的該等 J π — 71貝凋技術 標定出變化物體，並進行影像前處理工作，如去雜訊、區 域連！等，以便將變化物體之區塊找出，圖2⑴至2( γ)顯 示不範性前處理過程，苴分 〇刀另J為差值衫像、中值濾波、區 域連、，、D、頻譜能量過遽、 ^ ^ ^ 汉巴度貝成過濾。進一步地，針Figure 2(4) to 2 of the image generation process. In this embodiment, the images are the result of a typical fire simulated by the user, which was affected by the wind during the demonstration fire. In Fig. 2(g) and 2(h), there are obvious starting points for the fire, and Figure 2 (the image recorded by the wind blowing flame is recorded in the 。. The receiver's for the J π — 71 The Bayer technique calibrates the changing object and performs image pre-processing, such as de-noising, regional connection, etc., in order to find the block of the changing object. Figure 2(1) to 2( γ) show the non-standard pre-processing.苴分〇刀 Another J is the difference shirt image, median filter, regional connection,, D, spectrum energy over 遽, ^ ^ ^ Hanbadubei into filter. Further, the needle
對该變化物體進行「平 B r ^ , T θ 又值」及尚頻能量相對於全 頻此！的比值」之計算作為特徵向量。 、 最後可用如邏輯柵狀態的 ^ W乃式紀錄並儲存下如圖 所示特徵向量對時間的 物)中 Γ 数曲線21即為此事件發生時變 化物體的「平均g声佶 才欠 間的關係，而曲錄查L 士 玍旳h况）對時 处曰4 敦即為此事件發生時變化物體的「古相 旎篁相對於全頻能量 妁阿頻 值」（即相對於煙霧產生的情況） I15I38.doc 13〇7〇6] 對時間的關係。在此實施例中 ^. 了硯察到當疑似火災情況 發生時，此兩曲線均有如圖 人滑況 HI7A、、士， 守間點23(其相對於圖2(g)至 ^離的不連續跳躍現象產生；同樣 ) 消失時，會產生反向的跳躍情:只〜 田人焰次煙務 對於心、由"⑽況，例如圖中時間點24係相 :圖”風吹入造成火焰及煙霧的短暫消失。 §日後需進行監控火災事株β DM X事件疋否發生0寺，可針對連續輸 二同樣分析並即時榻取出監控區域範圍中變化物 體之「平均色度值及「 同頻尨I相對於全頻能量的比 並措由邏輯柵的運作狀態，判別是否 的ί災事件發生，並可進-步由系統發出警訊，達成= 功月"，或甚至可啟動自動灑水機制即時撲滅火災。前述之 邏輯柵狀態可改用習知的模糊邏輯柵狀態來描述， =性的射條件，使系統具有計算“發生火災，，之機率的 犯力’並依據系統内定或使用者自訂之臨界值來判定火災 事件的發生。 上述較佳實施例中雖為了簡化說明，僅選擇 ^為偵測事狀基準，且使黯僅提供—組包含該事^ 之&生過程之複數張影像作為預建模型用。但此技術領域 具通常知識者應能瞭解’本發明並不限於此，亦可自由.琴 取更多種特徵向量作為基準’錢供更多組影像來預建^ 型。隨者選取的特徵向量越多或預建模型時所提供的影像 組數愈多’則系統谓測出事件發生的準確度通常可相對提 问’准叶算複雜度亦相對提高。而使用者針對所欲哨測的 事件所選擇的特徵向量若對該事件的相闕性越高，亦可提 115I38.doc 16- 1307061 尚系統精確度，使用者 對應之特徵向量資料商㈣之常用_事件 用者自行嘗試驗證（自行;；=、/戈經由專家教育訓練、使 丁氣作杈擬影片測試）， 驗交換來提高所選擇的特 ’ 旧n # 量對該事件之適切性。 圖3展不根據本發明 — . 八體實施例。在此實施例 中’-種可僧測事件之影像 攝取裝置31，其在連嘖福童“充其包含：一影像 、“… 續複數個不同時間點攝取監看區33之 複數張影像，該影俊藉π ’攝取裝置31可為CCD或相 機、CCD或CMOS數位摄玛德斗、 仍數位相 办機或網路攝影機；及一八 析裝置32，其耦合至該影 〜刀 τ>像攝取裝置3 1以取得該耸 該影像分析裝置32可為個 …影像’ (DSP)處理器、或特定“ 處理器、數位信號The flat object is "flat B r ^ , T θ again value" and the frequency energy is relative to the full frequency! The calculation of the ratio is used as a feature vector. Finally, the W 乃 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 纪录 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑 逻辑Relationship, and the record of the record L 玍旳 玍旳 况 ) ) 对 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 敦 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化 变化Situation) I15I38.doc 13〇7〇6] The relationship to time. In this embodiment, it is observed that when the suspected fire situation occurs, the two curves have the same person's slip condition HI7A, shi, and custodial point 23 (which is different from Fig. 2(g) to ^ The phenomenon of continuous jumping occurs; the same) when it disappears, it will produce a reverse jump: only ~ Tianren flames for the heart, by the "(10) condition, such as the time point 24 in the figure: "The wind blows into the flame and smoke The short-term disappearance. § In the future, it is necessary to monitor the fire-fighting strain β DM X event or not. 0 Temple can be used for the same analysis of continuous transmission and take out the “average color value and the same frequency” of the changing objects in the monitoring area. The ratio of I relative to the full-frequency energy is determined by the operating state of the logic grid, and the occurrence of the λ disaster event occurs, and the system can send a warning to the system to reach = power month ", or even start automatic watering The mechanism immediately extinguishes the fire. The aforementioned logic grid state can be changed to the conventional fuzzy logic grid state to describe, the sexual shooting condition, so that the system has the calculation of "fire, probability of the force" and according to the system default or use Self-defined threshold To determine the fire incident. In the above preferred embodiment, for simplicity of description, only ^ is selected as the detection event reference, and 黯 is only provided as a pre-built model including a plurality of images of the & However, those skilled in the art should be able to understand that the present invention is not limited thereto, and that it is free to use a variety of feature vectors as a reference for money to provide more images for pre-built. The more feature vectors are selected or the more image groups are provided when the model is pre-built, then the accuracy of the system is usually relative to the problem. If the feature vector selected by the user for the event of the desired sentiment is higher in the event, the accuracy of the system may also be raised. 115I38.doc 16-1307061 is still systematically accurate, and the user corresponding to the feature vector data provider (4) is commonly used_ The event user tries to verify on his own (self;; =, / Ge through the expert education training, so that Ding as a virtual film test), test exchange to improve the relevance of the selected special 'old n # quantity. Figure 3 is not in accordance with the present invention - an eight-body embodiment. In this embodiment, the image capturing device 31 of the measurable event, in which Lianfu Futong "includes: an image, "... continues to ingest a plurality of images of the monitoring area 33 at different time points, The image capture device 31 can be a CCD or camera, a CCD or CMOS digital camera, a digital camera or a network camera; and an eight-input device 32 coupled to the image to the knife τ > Like the ingesting device 31 to obtain the image analyzing device 32, the image can be an image processor (DSP) processor or a specific processor or digital signal.
成特疋應用積體電路（ASI 階段，該影像分析裝置32可依據使用者所選擇的== 選項，針對使用者提供之_或數組紀錄有所欲制 複^張影像，擷取可代表該事件之發生的特徵向量時變模 型並加以儲存，以供日後 复模 後比對使用。在正常監測模式84 影像分析裝置32針對攝取裝^ 飞時， 丁攝取裝置31所即時拍攝之連續 行分析，判別其特徵向量時 象進 /里吋間函數疋否符合預存於 甲的特徵向量時變模型 庫 斷該事件的發生。右付合备度達到設U準，則列 本案說明書内文暨附圖應視為解說性的，而不 制性的。所有此類的修改皆屬本發明料内。 ‘、’、限 【圖式簡單說明】 種可 圖1⑷至l(q)展示根據本發明之__具體實施例之 I15138.doc 1307061 偵 偵 測-特殊行為發生之事㈣像監控方法及H 圖2⑷至2(δ)展㈣據本發明之—具體實施例之一種可 測火災發生事件之影像監控方法及系統。 圖3展示根據本發明之一具體實施例之 發生事件之及系統。 可偵測火 <« 【主要元件符號說明】 11 12 21 22 30 31 32 …阻」的時間函數 7骨架長短轴失角值」的時間函數 平均色度值」的時間函數 「高頻能量相對於全 θ 時間函數 頻此I的比值」的 影像監控系統 影像攝取裝置 影像分析震置 115138.doc -18-The integrated circuit is applied (in the ASI stage, the image analyzing device 32 can select the image to be provided by the user according to the == option selected by the user, and the image can be represented by the image. The feature vector time-varying model of the occurrence of the event is stored for later use in the complex mode. In the normal monitoring mode 84, the image analyzing device 32 performs the continuous line analysis of the instant shooting of the D-intake device 31 for the ingestion loading and unloading. When the eigenvector is discriminated, the eigen-in/in-time function is in accordance with the eigenvector pre-stored in A. The time-varying model library breaks the event. The right-handedness is set to U, and the text of the case is attached. The drawings are to be regarded as illustrative and not in a singular manner. All such modifications are within the scope of the present invention. ', ', limitation [schematic description] Figures 1 (4) to 1 (q) are shown in accordance with the present invention. __I15138.doc 1307061 of the specific embodiment Detecting detection - the occurrence of special behavior (4) image monitoring method and H Figure 2 (4) to 2 (δ) exhibition (four) according to the present invention - a measurable fire event Image monitoring method Fig. 3 shows a system for generating an event according to an embodiment of the present invention. Detectable fire <« [Description of main component symbols] 11 12 21 22 30 31 32 ... time function of 7 skeletal length axis The time function of the time function average chromaticity value of the deviation value "the ratio of the high frequency energy to the full θ time function frequency I" is the image monitoring system image acquisition device image analysis shock 115138.doc -18-