近年來，隨著無人飛行器（UAV）的生產技術不斷的進步，空拍影片的取得變得越來越容易，也因如此，針對空拍影片分析的研究，成為一個熱門的議題。而UAV拍攝影片的運用非常廣泛，舉凡軍事作戰、執法系統、搜救、以及交通的監控與管理等。 空拍影片的獲得，通常是將攝影機架設在飛機或UAV，由於此類飛行載具是在高空進行拍攝，因此影片具有視野廣、解析度低、拍攝高度多變的特性，此外，隨著飛行載具的移動，攝影機所拍攝的背景畫面也不斷地在改變，另因攝影機是架設在飛行載具上，攝影機也會有抖動、轉動、偏移等等現象，而地面上的車子本身也會移動，造成了空拍畫面中存在著許多的移動向量，也讓整個問題變得複雜許多。在本研究中，我們希望能對空拍影片中的車子，能有效的進行偵測與追蹤，我們運用TLD（Tracking – Learning - Detection）的架構，提出一個高效能的偵測與追蹤的方法。 為了能有效的偵測並追蹤空拍影片中的車子，一開始我們利用DBNs (Dynamic Bayesian Networks)先行偵測車子，接著使用SURF特徵點當作追蹤器進行追蹤。但是因為在空拍影片中，解析度通常是較低的，這也造成了特徵點會出現不穩定的現象，為了解決這個問題，我們運用蒙地卡羅的概念及補償的機制進行調整，並運用權重法進行特徵模型的更新，這樣的設計，讓追蹤的效果大幅的提升。另外，我們運用MHI (Motion History Images)及Corner Image的互補特性在車子的偵測上，實驗結果證明，這樣的方法能有效的改善DBNs偵測演算法存在的偵測錯誤問題。 ;With rapid advance of Unmanned Aerial Vehicle (UAV) design and manufacture technologies, the analysis of aerial videos taken from aerial vehicle has become an important issue. It has a variety of applications, such as military, law enforcement, search and rescue, and traffic monitoring and management. One of the most important topics in aerial surveillance system is vehicle detection and tracking. In this study, we propose a vehicle tracking system for aerial surveillance videos. The experimental videos in this work are with low frame rate, low resolution, and variable altitude. To achieve our goal, we utilize Dynamic Bayesian Networks (DBNs) to detect vehicles in the initial step. And perform an adaptive detection model via Motion History Images (MHI) and corner images to detect vehicles during detection iteration. We also extract feature points on the detected vehicles for tracking. In the aerial videos with low contrast and low resolution, the feature points are not stable in successive frames. In order to solve this problem, we design a weighting scheme and incorporate the concept of Monte Carlo methods to update the vehicle feature points. In our proposed method, we perform particle sampling around the vehicle feature points and acquire the appearance models of the image patch at the sampled points. The sampling area is dynamically adjusted according to the matching conditions of vehicle feature points. The vehicle feature points are updated by the sampling point with the highest similarity measure with target appearance model. The weights of the feature points are also updated to give the confidence level of tracked feature points. Experimental results have validated that the proposed method can substantially improve the detection and tracking accuracy on a challenging dataset.
