現今的監控系統中，電腦仍無法達到監控人員所能達到的判斷與分析能力。大部分的系統仍仰賴監控人員觀看監控畫面。由於日益龐大的需求，智慧型監控系統中相關的議題也受到廣泛的討論與重視。在智慧型監控系統當中，處理前景物追蹤所需面臨的遮蔽問題至今仍無法得到很好的解決。當多個前景物彼此遮蔽時，現存的追蹤演算法往往會得到不正確的追蹤結果。本計劃的特點在於希望能夠處理人群的活動分析。在本計劃中，首先針對追蹤彼此遮蔽的前景物提出改善正確性的方法。接下來，對於在畫面中無法分割成個體的人群，做人數估測的動作，以解決追蹤演算法先天上對於處理人群之限制。在完成追蹤和分析之後，即可針對所得的資訊瞭解並描述人群間互動的過程與監控影片之內容，並判斷特定事件之發生。在追蹤演算法的部份，計畫使用改進動態粒子採樣技術，分析物體的外型特徵以利正確追蹤。在群體人數估測方面，目標是不受攝影機擺設角度與物體大小之影響，需要先做物體大小之正規化與抽取不受視角與縮放影響之特徵，本計劃將研究設計適合於判斷群體中人數之特徵，選用發展適合之分類器。在人群活動描述與事件分析的部份，本計劃將研究並比較隱藏式馬可夫模型與條件隨機域作為辨識模型。 ; Currently, most surveillance systems still depend on surveillance personnel to monitor the surveillance scenes and computers are still unable to perform analysis and make decisions as humans can. Due to massive and growing demand, relevant topics in surveillance systems has been widely discussed and emphasized by researchers. The occlusion problems of moving object tracking in intelligent surveillance systems still cannot be entirely resolved until now. When more than one foreground objects occlude one another, existing tracking algorithms still can get erroneous results. The distinguishing feature of this project is analyzing group activities. Therefore, we would like to enhance the occlusion handling ability of the system. Then, for groups that cannot be divided into individuals in the surveillance scene, we would estimate the number of members in the group. Afterwards, the system can understand and describe the interaction between groups and detect event of interest. To improve the tracking algorithm and occlusion analysis, we plan to use enhanced dynamic particle sampling, and analyze the appearances of the objects for accurate tracking. In group estimation, our goal is to be invariant to camera angles and object sizes. Therefore, we will normalize the group objects and extract features that are invariant to viewing angle and scaling factors. We will select appropriate features and design suitable classifiers. In group analysis description and event detection, we plan to experiment with Hidden Markov Models and Conditional Random Fields and compare their performance. ; 研究期間 9711 ~ 9810
