In order to meet the needs of intelligent video surveillance, an unsupervised abnormal detecting algorithm was proposed. Firstly, model of mixture of Gaussians was used to extract the motion area, and the motion area was labeled. Then, observation sequence updated in real-time of feature matrix was established by the optical flow features obtained from labeled area which was normalized to the feature matrix. Finally, applying reconstruction works of two-dimensional principal component analysis on the sequence, abnormal behavior can be detected according to the energy ratio between the recovered feature matrix and original feature matrix. Experiments were conducted on various video datasets, which shows the effectiveness of the proposed method.