建立一套以電腦視覺為基礎的智慧型監控系統，在較少的監控人員下，讓系統自動地對攝影機所擷取到的影像進行分析，以減少人力和降低異常事件發生的傷害程度。 為了使監控系統具有判斷異常的功能，某些論文的方法是在特定的場景之中定義正常或異常的行為，而此行為並不適用於其它場景，因此必須根據各場景重新定義其行為特徵，因此本論文希望藉由場景中的物體，分析其正常的行為特徵，以此訓練完的結果作為異常判斷的準則。本論文提出一套異常行為的偵測系統，利用軌跡的特徵，來判斷進入場景中的物體是否行為異常，首先利用histogram的方法建立背景影像並使用HSV中的亮度資訊偵測是否有前景物後，將非前景物的陰影區域加以去除，並利用bounding box distance measure的方法來追蹤此前景物，接著將追蹤時所得到的軌跡資訊用正規化的vector來表示，輸入結合FSOM與改良式PCM的非監督式演算法作訓練後，以此訓練完的結果來判斷後續物體是否行為異常。 實驗結果證明背景模組和訓練軌跡的非監督式演算法，可有效地處理光線變化和訓練軌跡特徵，而此訓練的結果可正確且快速地偵測異常事件。 The developing of a computer-based monitoring system is an effective approach to monitor the space with less man power. An intelligent program could provide not only the recording function but also the prediction of abnormal activities. Trajectory feature was proven to be an effective feature for detecting the abnormal activities. However, the moving trajectories of objects should be pre-defined in traditional approaches. Since the monitoring scenes were varied widely, pre-defined trajectories are not available for all scenes. In this thesis, the video data with normal activities were collected and segmented to train an unsupervised learning model of normal behaviors. In this thesis, a fuzzy self-organized map (SOM) is built to detect the abnormal activities using the trajectory features. First of all, moving objects are detected and tracked in the histogram-based background subtraction, shadow removal, and labeling steps. The trajectory features of moving objects were extracted and represented as a normalized feature vector. The activity patterns are thus constructed using an unsupervised learning algorithm. Unlike the existing learning method, the proposed method combines the FSOM and the modified possibility c-means clustering algorithm. The parameters of SOM were replaced with the membership functions. They are repeatedly adjusted to obtain the desired output by the training samples. After completing the learning process, a normalized trajectory vector is classified to verify its validity. Experimental results are illustrated to demonstrate the effectiveness and efficiency of the proposed approach. The abnormal activities can be detected in a real-time video surveillance system. Finally, conclusions and future works are given.