- 學位論文
藉由弱監督式學習來偵測日常影片中寵物之異常行為
Weakly-Supervised Anomaly Detection for Pets in Daily Life
摘要
隨著現代人生活習性的改變，人與寵物關係已逐漸轉變為家人般的存在，而牠們的健康、環境及生活品質也成為人們關切且重視的議題。近來寵物攝影機盛行，方便飼主隨時監測寵物的行為狀況；然而現況是，飼主往往在意外發生時無法及時發現，需要事後透過回放攝影畫面來檢視。若能使用異常偵測 (Anomaly Detection)將端到端模型應用於偵測寵物行為，在異常行為發生時及時發現及應對，相信將能大大地減少傷害。 由於目前沒有寵物相關的異常偵測資料集，且有鑑於家用寵物攝影機畫面取得不易且多半涉及隱私，本研究轉而蒐集以貓為對象的手持相機拍攝影片，創建了寵物異常資料集(CATA)，包含咳嗽、癲癇和嘔吐等貓的常見異常行為；同時搭配一套資料前處理流程，將影片經過場景切換偵測及篩選、去除不適合的片段，並進行影片畫面穩定化處理，使手持裝置拍攝的影片得以如監視器畫面般穩定，以利後續偵測。 基於異常影片蒐集不易，加上影格標註需高額人力成本，我們透過弱監督式學習結合多示例學習來訓練模型於只標註異常與否的影片上以降低成本；運用時序注意力模塊 (Temporal Attention Module) 找出包含異常影像的關鍵影格和分群模塊 (Cluster Module) 以區分異常與正常影格。最後，我們修改了原用於異常偵測的損失函數，使其更符合CATA 資料集，進而得到模型在預測寵物異常行為上的出色成果。
並列摘要
With the changes in the living habits of modern people, the relationship between people and pets has gradually transformed into a family; their health, environment, and quality of life have also become issues of concern and importance to people. Recently, pet cameras have become popular, which is convenient for owners to monitor the behavior of their pets at any time; however, the current situation is that owners often cannot find out in time when accidents happen and need to review the camera images afterward. In this regard, we utilize Anomaly Detection to apply an end-to-end model to detect pet behavior, so as to facilitate timely detection and response when abnormal behavior occurs. Since there is currently no pet-related anomaly detection data set, and considering the difficulty of obtaining pet camera videos which involve privacy mostly, this research turned to collecting videos taken by handheld cameras with cats as objects, and created a cat anomaly detection dataset (CATA), including common abnormal behaviors of cats such as coughing, seizure, and vomiting. In addition, data preprocessing is established to process the video through shot change detection and selection, unsuitable clip abandonment, and video stabilization so that the video shot by the handheld device can be as stable as the surveillance camera for subsequent detection. Due to the difficulty in collecting anomaly videos and the high labor costs of frame labeling, we use weakly-supervised learning combining multi-instance learning to train the model on videos that only label abnormal or not to reduce costs; leveraging the temporal attention module to find the key instance containing abnormal events and the cluster module to distinguish abnormal and normal frames. Finally, we adapted the loss function originally used for anomaly detection to make it more consistent with the CATA dataset, resulting in outstanding performance for the model in predicting abnormal behaviors of pets.