於複雜環境下以適應性學習演算法進行攝影機異常之偵測

Abstract

當犯罪事件發生前，經常發現視訊監控攝影機遭到人為蓄意破壞，使得攝影機無法有效的錄下犯罪事件發生經過。本論文提出一個自動且即時的攝影機異常偵測方法，能在複雜的環境下，偵測出是否有人為蓄意破壞，並能有效降低環境變化產生的誤報機率。本方法採用二階式偵測，第一階段以取樣點方式代替整張影像作為差異比較與分析，可以明顯提升執行速度。並以適應性學習方式，使得取樣點能夠穩定且均勻分布於場景的邊緣上。藉由分析取樣點的灰階強度變化情形，判斷是否有異常事件發生。第二階段開發誤報事件偵測器，藉此過濾誤報事件，以降低誤報率。本研究針對實務上極易發生之開關燈誤報事件，提出一個有效的解決方法。此方法為利用影像結構相符程度來判斷該警報是否來自於光線變化所產生，而非真正的攝影機異常。由實驗結果，本方法能夠有效且即時偵測出攝影機遭到遮蔽、模糊和轉向等異常，且對於正常的環境光線變化、大型物件及大量人群經過等，皆比過去的方法不易產生誤報。

When a crime event happens, criminals often tamper the camera to prevent their suspicious activities being captured. In this thesis, we propose a method to detect camera tampering, which can run in real-time and can be applied to a complicated environment. Our method includes a two-stage detection framework. In the first stage, we use edge intensity as the main cue to detect the event of camera tampering. Instead of using all the edge points in the image, we sample some edge points to speed up the system. In addition, we propose a learning method to determine the sampled points, and it guarantees the sampled points would be stable and uniformly distributed in the image. According to analyzing the variation of edge intensities of the sampled points, the event of camera tampering can be detected. To reduce false alarms, the second stage is trigger when the first stage detects the event. The second stage is to check whether the events come from false alarm or not. In this thesis, an illumination change detector is proposed to check and reduce false alarm. It matches the edge blocks to determinate whether the tampering event comes from the illumination change or not. In the experiments, we demonstrate the proposed method can detect the camera tampering well and can avoid triggering false alarm even when the illumination changes dramatically or large crowd passes the scene.