利用前景運動資訊之異常行為定位偵測

Abstract

監視攝影機系統現在已經廣泛使用在全世界。本研究藉由電腦強大的運算能 力，對一段影像做一個初始的處理－也稱為冷開始－，找出影片當中使用者，例 如警方可能會有興趣的片段藉以減少人力資源耗費。本研究提出了以物體當作基 礎，再以光學流動來推測前後幾張畫面物體的運動趨勢，就物體時間以及空間上 面的關係進行預測，並且以物體的運動行為當作特性，以機率模型加以描述，再 分析不同物體之間的差異性，以找出影片都中可疑的片段。經過實驗之後，證明 本研究在異常行為定位上面有優異的表現。此外，本研究在未來有許多實際的應 用，例如偵測搶案、丟棄物品以及行人跌倒等異常行為。

Surveillance system is pervasive all over the world. In our research, We exploit powerful computation intelligence to preprocess the video clips, also called cold-start, to filter out the clips which users, the police for example, might be interested in, in order to reduce the human resource cost. Object-based method is proposed, and optical flow is used to estimate the motion tendency in consecutive frames, to increase the spatio-temporal relations. We use motion as features, which are described using probability model. And we analyze the differences between blobs to filter out the suspicious clips in the video. Through experiments, our method has excellent performance on anomaly detection localization. Besides, in the future, some practical applications are expected, such as detection of robbery, abandoned objects and human-falling-down.