隨著電腦科學的進步和對社會安全的迫切需要，以視覺為基礎的非監督式監控系統成為近年來熱門的研究主題。所發展的系統可應用在許多不同的領域上，例如保全服務和交通監控。而使用新系統具有許多好處，包括節省人力資源、降低成本和提供一致性的服務。 對交通監控而言，能正確的偵測交通違規事件扮演非常重要的角色。本篇論文提出一個新的方法能針對特定車道(特別是內側車道)的大型車輛進行偵測和追蹤。在所提出的方法中，首先我們產生一張activity map進行車道偵測，並計算出有用的資料包括車道寬度和消失點以利後續的工作。接下來偵測模組利用連續影像相減和Sobel邊緣偵測在偵測區域找出大型車輛。在追蹤步驟則是採用卡曼濾波器來進行追蹤工作。在這裡我們推導出一個時變的狀態轉換矩陣以適應速度在2-D影像上的變化。再者，為了使追蹤更有效率，我們採用雙模式的追蹤模組。 實驗部分是採用數段不同的實際交通影像，大型車的偵測與追蹤平均準確率分別為91.3%和84.4%，實驗結果顯示論文所提出的方法能準確且有效率的偵測並追蹤大型車。 With the advancement of computer technologies and the urgent demand for social security, researches on vision-based surveillance grow more and more important. The developed systems can be employed in various applications, such as security service and traffic monitoring. The advantages of using such systems include the saving of human resources, the reducing of costs, and the providing of consistent performance. The correct detection of traffic violation events plays a very important role in traffic surveillance. In this thesis, a novel approach is presented to detect and track large vehicles driving on specific lanes (especially the inner lane). In the proposed approach, activity map is firstly generated to detect lanes, and useful data is extracted including the lane width and the vanishing point to facilitate the later task. Secondly, vehicle detector is devised to find large vehicles in the detection area by utilizing the techniques of temporal difference and Sobel edge detection. In the tracking process, Kalman filter is adopted to accomplish the task. Here, a time-varying state transition matrix is devised to adapt the velocity variations in 2-D images. Moreover, dual mode tracker is developed for more effective tracking. Experiments were conducted on a variety of real world traffic scenes. The average accuracy rates of large vehicle detection and tracking are 91.3% and 84.4%, respectively. Experimental results reveal that the proposed approach is feasible and effective for large vehicle detection and tracking.
