Abnormal event detection in surveillance videos has been a popular topic in computer vision area.We propose abnormal event detection by learning bases of event Attributes.Attributes are defined as the typical features of the learning data like words in a dictionary.We model the attributes by learning a set of feature bases.Then,an image patch in video can be represented by the attributes.A normal image patch can be reconstructed by the linear combination of the attributes and the reconstruction coefficient is sparse.However,an abnormal event image patch has high reconstruction residual or low reconstruction residual but dense coefficient.Experiments are conducted on UCDS datasets.Results show that the proposed approach could reliably locate the abnormal events in the video sequence,outperforming the current state-of-the-art methods.