- 學位論文
長時序動態影像之前景分割與巨觀模式下之異常行為偵測
Foreground segmentation with multiple temporal images and macro view-based rare event detection
摘要
在安全監控的各種應用之中，現有之行為偵測方法大多採用微觀的角度進行分析，必須在前景主體完整的輪廓下，找出每個肢體的位置，並事先定義各種動作與姿勢，但在固定攝影機所監控的實際環境中，通常面臨光源變化、背景物體移動等狀況，以及行為可能由單人或多人所構成，種類相當眾多且複雜，無法逐一定義，更無法預先知道異常行為的種類，因此本研究利用機器視覺之技術，將環境中的移動前景主體有效分割出來，並且透過巨觀之策略進行異常行為之偵測，不需事先定義各種行為或動作。 本研究主要兩個目標為：1)開發有效之動態前景分割技術，和2)視訊影像中之異常行為偵測。在前景分割方面，首先透過多張時序影像所構成之灰階紋路特性提出兩種前景分割方法，傅立葉轉換為基礎之前景分割方法能有效濾除屬於靜態背景之紋路，此方法具有完整的輪廓以及不受背景改變的干擾，但運算時間較長；統計管制界限為基礎之前景分割方法依照多張時序影像之平均值與變異數變化情形，將超出管制界限之前景分割出來，具有偵測靜止不動前景的能力以及不受背景改變的干擾，且計算快速。在異常行為偵測部分，本研究針對前景分割結果中移動主體在時間與空間維度所造成的動態變化建構出全域式的表達方式，可同時紀錄每個主體的行為與移動情形，將各種行為表達成一張全域式的動態能量圖，透過此策略可免除現有行為偵測需進行前景肢體分割的困擾，且全域式的表達方法不需事先定義行為的種類與持續時間；收集日常生活之各種正常行為模式之後，透過模糊C-means進行正常行為之群集訓練，以作為異常行為之偵測標準。 本研究以實驗室的日常活動作為探討與驗證對象，收集兩天的正常行為樣本並建立出正常行為之群集，目前測試暈倒、小偷、打架以及搬家四種異常行為，皆可有效偵測出來，並且長時間觀察實驗室31天之實驗結果中發現異常行為偵測之方法能夠穩定且有效找出以上四種不同於日常生活中的異常行為，在影像尺寸200x150的實驗中皆能達到即時偵測的效果。
並列摘要
In applications of video surveillance, good foreground segmentation ensures the success in subsequent activity recognition. The traditional behavior recognition in video images is based on micro-view analysis. It must separate a moving body into detailed parts of head, hands, trunk and legs, and the state changes of the parts are then estimated for behavior classification. The success of this approach highly relies on precise segmentation of the full shapes of moving objects. This research proposes machine vision algorithms for fast/precise segmentation of foreground objects, and studies the feasibility of rare event detection from a macro observation. This study has two main goals: 1) create efficient foreground segmentation, and 2) detect rare event in video images. In foreground segmentation, this study proposes two algorithms based on gray-level structures present in multiple temporal images. This first algorithm uses the Fourier transform to remove the stationary background and preserves the complete shape of the foreground. The second algorithm applies statistical process control to separate foreground with fast update of gray-level mean and variance of multiple temporal images. In rare event detection, this study proposes a macro-view representation that records the time-space motion energy of all moving objects in a scene without segmentation of individual objects and their details parts. By collecting normal daily behaviors and using a hierarchical fuzzy C-means to cluster normal behaviors, the observed event that has distinct distance from any of the cluster centroids is classified as an anomaly. This approach prevents the difficult definition of various types of rare events. This study took daily living in a laboratory for experiment. Four rare events including fainting, burglary, fighting, and moving furniture out of the laboratory were evaluated. The detection results of a long sequence of observations over a month have shown the efficacy of the proposed method for on-line, real-time monitoring of video images of 200x150 size.