多視點視訊產生、傳輸與分析-子計畫五：固定式與移動式多攝影機交通影片內容理解、分析與管理

Abstract

由於硬體價格降低，攝影機越來越普遍地架設在平面道路、高速公路以及私人車輛之上，但這些攝影機的資訊，只有在事故已經發生或是需要追查通緝犯時才會由人主動調閱查看，往往需要花費許多時間來看影片。因此許多研究致力於影片內容理解與分析來研發實用的工具系統，讓使用者可以快速且有效的獲得所要的影片資訊。目前多數研究著重於固定式攝影機的整合系統與單一移動式攝影機之影片分析，然而固定式攝影機由於經費限制，即使整合所有固定式攝影機，監控涵蓋率依然不高，只能針對重點路段監控。而單一移動式攝影機之影片分析受限於單一視角，只能透過對應點推論得到較不準確的三維資訊。因此，本計畫的研究目的除了使用影像低階特徵值來做物件切割、軌跡擷取與事件偵測外，更使用多重視角資訊來增強影片分析，並整合固定式與移動式多攝影機資訊，提升影片內容分析理解的準確率與涵蓋率。 第一年我們研究如何從固定式與移動式影片中準確地切割移動物體並擷取軌跡，且提出以物件軌跡為基礎之事件偵測之演算法，除此之外，我們研究近似複製影片檢索來進行影片管理。第二年利用多視角技術來做影像分析，並將第一年之各式演算法修改使其適用於夜間、市區等較複雜環境，且利用軌跡與事件資訊來做近似複製影片檢索。第三年我們考慮各種不同天氣狀態，並提出適用於不同天氣狀態之物件切割、軌跡擷取與事件偵測演算法，並整合移動式與固定式多攝影機資訊，建置各式交通影片內容理解、分析與管理系統。

With the affordable prices of digital capturing devices, more and more cameras are installed on plane road, highway and personal vehicles for traffic surveillance and event recording. Generally, people view the recorded videos manually when events happened, and the viewing process is a boring and time-consuming task. To provide effective and efficient video retrieval systems, many researchers put their eyes on video content understanding, analysis and indexing. The majority of the existing related works focus on two aspects: the integration of static multi-cameras video contents, and second, single-view mobile camera video content analysis. However, due to the limit of budgets, the cover rate of static cameras is not enough for traffic surveillance. Furthermore, single-view mobile video analysis is limited by incomplete information. To deal with the above problems, we use multi-view information for mobile video analysis, and integrate the static and mobile multi-camera video information to improve the effectiveness and cover rate of intelligent transportation systems. First year, we focus on designing the algorithms for object segmentation, trajectory extraction, and trajectory-based event detection for both static and mobile daytime traffic videos. Besides, the near-duplicate video retrieval technique is developed for video database management. We further extend the algorithms developed in first year to complicated situations, such as nighttime and urban traffic videos in the second year. In addition, we apply the multi-view information for mobile video analysis to provide effective driver assistant systems. The trajectory- and event-based near-duplicate video retrieval is implemented for traffic video management. For the third year, we take various weather conditions into account to propose video analysis methods. Finally, we integrate the static and mobile multi-camera information to build traffic video understanding, analysis and management systems with high accuracy and high cover rate.