同調性事件壓縮之影像縮時

Abstract

監視攝影機目前被廣泛地使用來監控生活中各式各樣的事件，數以萬計的監視紀錄使得後續的瀏覽以及追蹤異常事件變得十分耗時。過去研究為了提升瀏覽影片的效率而提出將影片中沒有事件發生的時空區塊移除，然而這樣將畫面呈現最大化的利用反而容易分散使用者的注意力。在注意力的相關文獻中顯示，當使用者必須同時追蹤影片中不同區塊的多重事件時，使用者反而容易出錯並且需要更多時間來消化影片中的資訊。本研究提出一種同調性的影像縮時方法來解決上述的缺點，採用事件發生的軌跡來做為同調性的依據具有以下兩個優點: (1) 易於檢視多個相似軌跡事件同時呈現的畫面;(2) 異常事件的移動軌跡通常異於常態事件。 本研究使用最長共同子序列演算法做為判斷不同軌跡間相似度的方法，此方法著重於軌跡間雷同的部分，因此對於不同的部分具有較高的彈性，具有抗雜訊和比對伸縮軌跡的優點。實驗結果顯示所提出之同調性事件壓縮之影像縮時可得到滿意的呈現效果和壓縮效率。

In order to monitor all kinds of events in daily life, surveillance cameras are widely used nowadays. Millions of surveillance records present challenges to their browsing and retrieving. Previous approaches are proposed to address this problem by eliminating the inherent spatiotemporal redundancy as much as possible. However, maximizing the visual information in the synopsis frames may distract the visual attention of viewers. While simultaneously showing several events on one screen, users would make more mistakes or perform their tasks more slowly. Therefore, a novel synopsis system capable of making coherent video summary is proposed in this thesis to overcome the above challenges. Two main contributions of making the coherent synopsis video with trajectories are: 1) Events with similar trajectories are easy to retrieve simultaneously. 2) An abnormal event can be distinguished from normal events by analyzing moving patterns. Longest Common Subsequence (LCS) is applied to measure the similarity among trajectories due to its robustness and stretchiness. Finally, the clear display and efficient condensation can be obtained from the coherent event based synopsis video. Comprehensive experiments conducted on various surveillance videos demonstrate the satisfactory performance of our proposed system.