擁有能自動判斷一個場景中是否發生異常事件,或者能預測潛在異常行為的能力對視訊監控系統是很重要的。目前的異常事件偵測方法只依據物件的移動軌跡,或者只根據場景中整體內容的變化作為事件偵測的描述。以軌跡特徵為基礎的事件偵測方法受限於其必須依賴可靠的物件追蹤方法,而以畫面特徵為基礎的方法的缺點是以整體方式計算出來的特徵其辨識力可能不足以識別某些事件。在此研究中,我們提出一個結合軌跡特徵與畫面特徵,用於視訊監控系統異常事件偵測的架構。此架構使用的特徵集合包含豐富與穩定的資訊,可以有效的描述視訊片段中的動作與事件。我們將測試本方法用於常見的視訊監控應用,例如大廈保全或交通監控的效能。
Current approaches for abnormal event detection in video surveillance either based solely on object trajectories, or seek global changes in scene content as representations for detection. A limitation of trajectory-based approaches is that they depend on the existence of reliable methods for tracking moving objects, and the drawback of frame-based methods is that feature signal computed globally might not be discriminative enough to identify certain events. In this study, we propose a framework for abnormality detection using both descriptive trajectory features and robust frame features. The aggregate feature set contains rich and stable information for describing motion events in a video segment. We show the performance of the proposed framework on common video surveillance applications.