| 摘要: | 自動化安全監視系統需要被發展，不僅可以節省人力成本，而且對於安全環境的監控是一項重大提升。我們將討論三個基於攝影機在各種環境中的監視應用議題：異常人體姿態偵測、多相機聯合自動偵測與追蹤、及多旋翼機自主追蹤。以下就這三個議題各別闡述。
在居家環境中，老年或行動不便人員可能因疾病或功能障礙而在移動中跌倒，蹲下或倒坐在地面上。如果無法及時發現，則可能導致致命危險。為了減少致命事件，提出了針對老年或行動不便人員的環境監測系統研究。本研究提出的系統，考量了四種異常行為的偵測：(1) 跌倒與蹲下；(2) 未經允許出入房間；(3) 未經允許離開指定區域；及 (4) 記錄步行軌跡。系統使用全方位攝影機 (omni-directional camera) 進行監視及影像擷取。首先應用背景減法 (background subtraction) 偵測前景物，然後使用主成份分析法 (principal component analysis, PCA) 為基準及以梯形區域高度變化為基準的偵測方法來偵測跌倒和蹲伏姿勢。在所提出的方法基礎上，進行了多種環境下的實驗與評估。由實驗結果顯示，以主成份分析為基準及以梯形區域高度變化為基準的人員跌倒與蹲下偵測率分別為93% 與92%；病人擅離病房的偵測率為95%；而病人擅離病床偵測率為95%；系統在實驗結果皆能獲得穩定的偵測結果，證明了該系統的可行性。
我們所提的異常人體姿態偵測系統具有下列特色：(1) 使用全方位攝影機，能拍攝到360度的環場影像。(2) 使用主成份分析法偵測人體主軸的方向與長度。(3) 使用人體梯形區域高度變化判斷人員狀態。
單鏡頭攝影機因受到鏡頭結構的限制，其監看範圍是有限的。多魚眼鏡頭攝影機可以監視大範圍並追蹤移動物體的完整軌跡。本研究使用兩部魚眼攝影機，提出一個自動偵測及追蹤的環境監視系統 (surveillance system)；提出的系統由兩個主要模組組成：前景物偵測與前景物追蹤。首先應用背景相減法 (background subtraction) 偵測前景物。在前景物追蹤，使用卡爾曼濾波器 (Kalman filter) 預測前景物的位置，在相機重疊畫面預先建立轉換表，當前景物跨越相機監視範圍時，利用轉換表能了解不同攝影機追蹤的物件實際是同一前景物，並維持物件編號聯合追蹤。為了要增加比對的可靠性，加入前景物的外型特徵輔助判斷；例如，顏色。在系統運作時兩部攝影機皆可實現多目標的追蹤，並且持有各自的追蹤器，當物件發生短暫遮蔽仍持續預測位置並提供平順的移動軌跡。由實驗結果顯示，本研究使用兩部魚眼相機，並透過具有不同光線亮度和人物數量的多部影片做實驗，平均敏感度為96.7 %，誤判率為0.45%，判斷錯誤發生的原因是前景物的衣物和背景色彩很相似，當系統加入Kalman filter追蹤輔助後可以使敏感度提升為98.55 %。實驗結果證明系統能適應許多室內不同監視狀況，如光線變化、陰影干擾、遮蔽，皆能有效且穩定的聯合前景偵測和追蹤系統。
我們所提的多相機聯合自動偵測與追蹤系統具有下列特色：(1) 可適應光線變化和雜訊的前景偵測。(2) H-S色彩分佈圖比對及距離相似度比對。(3) 跨相機座標轉換，可連續追蹤目標物。
多軸旋翼機發展出被動式追蹤，也就是在被追蹤目標上放一個被追蹤裝置，多軸飛行器透過該裝置的訊號來跟蹤，但追蹤非特定目標時就無法事先放置被追蹤裝置在目標上，將造成追蹤失敗。本研究提出多軸旋翼機電腦視覺追蹤系統，透過電腦視覺達到追蹤任意目標而不需額外被追蹤裝置的目的。首先利用尺度候選圖改進KCF (kernelized correlation filters)，使KCF可以適應目標變化大小，接著使用基於特徵匹配的偵測演算法讓目標丟失時也能重新找回。在所提出的方法基礎上，進行了多種環境下的實驗與評估。本研究使用縮放和遮蔽兩種影片場景，做為測試演算法對於縮放與遮蔽場景的追蹤能力判斷；除此之外，我們利用追蹤率、重疊準確率和執行速度做為評估依據，首先使用人工方式框出目標在影像中的位置做為成功追蹤的判定依據，接著利用所框出的位置計算演算法的追蹤率、重疊準確率和執行速度。由實驗結果顯示，我們提出的演算法執行速度為26 fps、在縮放的案例中追蹤率為88%、在遮蔽的案例中追蹤率為 98%、重疊準確度為 87%，在速度與準確率的取捨中，我們犧牲了演算法執行速度來換取更穩健的追蹤。系統在實驗結果皆能獲得穩定的偵測結果，證明了該系統的可行性。
我們所提的多旋翼機自主追蹤系統具有下列特色：(1) 透過電腦視覺達到追蹤任意目標，不需額外被追蹤裝置。(2) 透過尺度候選圖改進KCF，使KCF可以適應目標大小變化。(3) 當目標遮蔽或目標消失時，利用基於特徵匹配演算法重新找回目標。
;Automated security surveillance systems need to develop, not only can save money on labor costs, but also for monitoring by the security environment is a major upgrade. In this dissertation, we discuss three application issues of camera-based surveillance in various environments: abnormal posture detection, automatic detection and tracking and multicopter autonomous tracking. These three research issues are respectively discussed as follows.
In various general environments, the aged or disabled people may fall, squat, or sit down on ground due to sick or dysfunction in move. If the events are unable discovered in time, the fatal danger may be then caused. To reduce the fatal events, an environmental monitoring system for the aged or disabled people is proposed. In this study, four kinds of abnormal behavior detection are considered in the proposed system: (i) fall down and crouch, (ii) going out a room without permission, (iii) leaving specified area without permission, and (iv) recording walk trajectory. An omni-directional camera is used to capture images for monitoring. The background subtraction method is first applied to extract targets. Then principal component analysis (PCA) and the change of height of a personal trapezoidal bounding box are used to detect fall down and crouch postures. Based on the proposed method, the experiments and evaluations in a variety of environments have been carried out. The experiments results show that the PCA method can perform sixty frames per second, and the height of the patient trapezoidal bounding box method can perform eighty frames per second. The detection rate of the fall down and crouch detection using PCA is 93%, and the detection rate of that using the height of the patient trapezoidal bounding box is 92%; the detection rate of leaving ward is 95%; and the detection rate of leaving bed is 95%. Stable detection results were obtained to show the feasibility of the proposed system.
The proposed abnormal posture detection method has the following properties: (i) using an omni-directional camera can easily take 360-degree surround images, (ii) using principal component analysis (PCA) to detect the main direction and length of the personal body, (iii) using the change of the height of the patient trapezoidal bounding box and then determine the personal situation.
The view scope of a single camera is ?nite and limited by scene structures. Multi-fisheye cameras can monitor a wide area and trace a complete trajectory of a moving object. In this study, an automatic detection and tracking system with two fisheye cameras for environment surveillance is proposed. The proposed system is composed of two major modules: foreground detection and foreground tracking. The background subtraction method is first applied to extract targets. Then use Kalman filtering for pedestrian motion prediction. A transform table is pre-established to associate multi-cameras data in the overlapping areas. When object across disjoint camera views, the data in the lookup table can provide enough information to realize the moving object in camera views actually belonging to the same object, and keep consistent labels on the object. To improve the reliability of the tracking performance, motion and color appearance features are used to match the detected objects in different cameras. The experiments results show that the average sensitivity is 96.7% and the average false positive rate is 0.45% because the foreground objects are similar to the background. The average sensitivity rises to 98.55% with the Kalman filter. It demonstrates that the proposed method can work well under challenging conditions, such as light change, shadow interference, object occlusion.
The proposed automatic detection and tracking method has the following properties: (i) proposed foreground detection and tracking method can adapt to brightness changes and noise, (ii) using H-S color distribution map comparison and distance similarity comparison to detection, (iii) using the camera overlapping features, multi-cameras coordinate conversion, continuous tracking of objects.
A multicopter is equipped by a passive tracking device to follow a specified target. However, if want to track a non-controlled target, the passive tracking device is failed. We propose a vision-based tracking system for multicopters, system used computer vision method to track any target without additional tracking devices. In this study, propose scale candidate graphs and scale tables to improve kernelized correlation filters (KCF). There are also stable results when the scale changes. The proposed an adaptable scaled KCF algorithm, when the KCF tracking failed, a feature-based matching detector is used to re-detect the target. In experiments, there are two cases of scaling and occluded targets in the videos to be evaluated. The experimental results reveal that the proposal method can adapt the change of scale and occlusion of targets. In addition to the above ability, the tracking rate, accuracy, and execution speed are also evaluated. The experiments results show that the proposed algorithm performs speed to 26 fps, tracking rate in the scaled cases can reach 88%, tracking rate in occluded cases can reach 98%, and overlap rate can reach 87%. With trade-off between speed and accuracy, we sacrifice the execution speed of the algorithm to exchange for more robust tracking. Several experiments on various scene based on the proposed approach were conducted and evaluated.
The proposed an adaptable scaled KCF method has the following properties: (i) allows multi-axis aircraft to have the ability to track non-specific targets through computer vision tracking, (ii) using adaptable scaled KCF can adapt to the target change size and get the correct target image, (iii) using matching points to find the target position and scale to successfully retrieve the target. |
