監控視訊影像之軌跡式物件搜尋

Abstract

本論文提出監控影片在經過前景偵測後，不利用前景影像內容的色彩、紋理與輪廓資訊而僅使用前景物體的外接矩形資訊，透過簡單地分析畫面間矩形的重疊程度來達到快速追蹤前景物體與建立軌跡之目的。完成後的軌跡再與追蹤過程中得到的軌跡特徵資訊整合，便可建構監控影片資料庫的軌跡式索引。而追蹤過程中，前景物體的外接矩形尺寸可能在某些因素的影響而產生變化，因此我們亦提出尺寸分析的方法，將具有尺寸變化的外接矩形進行調整，使其重疊程度之估算能更為準確。此外，我們亦可利用此外接矩形尺寸的分析方法來將交會中的前景物體個別分離出來。最後，本論文提供一個使用者搜尋介面，讓使用者可以輸入搜尋目標的移動方向或軌跡路徑來對監控影片的內容進行搜尋，透過與影片的軌跡式索引資料分析比對，便可以快速地回傳適切的影像片段，使搜尋的工作能更有效率。

In this thesis, we propose a tracking method that only employs bounding boxes of detected objects without using color, texture and contour information. With a simple overlap analysis of its bounding boxes in successive images, we can track an object and construct its trajectory very quickly. Using these trajectories along with their associated attributes obtained from object tracking, the trajectory-based indexes of a surveillance video database can be constructed. During the object tracking, undesirable size change of bounding boxes may occur. A novel size analysis method is developed to resize such bounding boxes to improve the results of overlap analysis. Besides, we can also employ the proposed size analysis of bounding boxes to help the separation of occluded objects. Finally, a search interface is provided, so that a user can search a target by specifying a direction or sketching a trajectory. By using the trajectory-based indexes, the search interface will only return video segments with proper trajectory attributes to users, making the search effective and very efficient.