隨著近年來炸彈恐怖攻擊的增加,而且恐怖攻擊大部分都發生在人潮密集的公共場所,所以如何設計一個偵測靜止可疑(遺棄)物體的監視系統是非常重要的;在本篇論文提出了一個監視系統的技術,希望能在前景(foreground)較複雜的公共場合裡,儘快且不失準確性地找尋出靜止不動的可疑物體。 有一些研究也是在討論這類的問題。他們大部分都是利用靜止不動的可疑物體的兩個特性來設計其監視系統: a)靜止可疑物體不屬於背景。b)以及在監視的影片中,靜止可疑物體出現在頻率會比其他的前景出現的頻率高,或是靜止的可疑物體在某個時間內會連續的出現在影片中。 上述的監視系統在人潮並沒有十分擁擠的環境中,可以得到不錯的監視效果,但是當系統用來監視人潮非常擁擠場所時(如在捷運台北車站),由於靜止物體可能大部分時間都被行人給擋住了,導致上述的靜止物體的第二特性就不明顯,這些方法就不適用。所以此篇論文主要是設計一個動態過濾器(motion filter),先將每張frame中極有可能是動態物體的點先移除掉後,紀錄剩下可能為靜態物體的點。倘若某個點,在video sequence中出現靜態物體的點的次數累積足夠多,且顏色均相似,我們才宣告此點為可疑物體上的點。而我們套用現在model background很熱門的方法之一-Mixture of Gaussian,來紀錄的每個點顏色的歷史紀錄。 最後,我們針對三種人潮量不同的video sequences來做實驗。我們發現在人潮量最少以及人潮量普通的環境裡,沒有使用我們設計的動態過濾器,也可以找出影片中的靜止物體;但是當人潮非常眾多的時候,倘若沒有使用我們的動態過濾器,會出現許多的錯誤警告。如果加上我們的動態過濾器,就可以移除掉這些來自於人潮造成的干擾(和其他方法比較,約減少70%的錯誤警告)。
As the increasing of the bomb attacks in recent years and these attacks are repeatedly concentrated on the public places, such as MRT stations. Establishing a surveillance system with high-tech appliances to against terrorism becomes a critical issue nowadays. In this thesis, an algorithm of finding the abandoned objects in the environment of crowded public places is proposed. There exist some approaches to discover the abandoned objects under the circumstance of two scenarios: 1) The pixels in the scenes of abandoned objects do not intermixed with background pixels; 2) The abandoned objects emerge more often than other moving objects (pedestrians, commuters, etc.) in a surveillance video segment. These approaches work well in the phenomena of occasionally few pedestrians, but may fail in case there are crowded in rush hours; the system would issue many false alarms then. In this study a motion filter is proposed to filter out the partial scenes caused by irrelevant motion objects, together with accumulate the remaining useful pixel information. If we have enough evidence of abandoned objects according to cumulated records, an alarm is issued. We use the Mixture of Gaussian (MoG), which is the most popular background modeling tool, to record the useful history of each pixel. Finally, we use three scenarios to examine the performance of our approach, they are: few, normal and rush hours. The detection accuracy of the system without our motion filter is still satisfying in the environment with casual or normal cases, however it issued numerous false alarms in the environment that is highly crowded. While on the contrary the system with the help of our motion filter will issue the proper alarm for abandoned objects or some few unfavorable alarms for some specific background noises.