Unsupervised learning approach for abnormal event detection in surveillance video by revealing infrequent patterns

Abstract

Uncovering the hidden subtleties and irregularities of the events in the video sequence, is the key issue for automatic video surveillance. Notice the fact that the occurrence of abnormal events is rare while the frequently occurring events become normal in general human perception. So we have proposed the unsupervised learning algorithm, Proximity (Prx) clustering for abnormality detection in the video sequence. Prx clustering tries to select only the dominant class sample points from the dataset. For each data sample, it also assigns the degree of belongingness to the dominant cluster. The proposed motion features viz. circulation, motion homogeneity, motion orientation and stationarity try to extract important information necessary for abnormality detection. After performing Prx clustering, each sample belongs to dominant cluster with the membership value. When Prx clustering is being performed in the space constructed from the proposed motion features, it helps to improve the abnormality detection performance. Experimental results for clustering performance evaluation on artificial dataset show that the Prx clustering outperforms the other clustering methods, for clustering the single dominant class from the dataset. Abnormality detection experiments show the comparable performance with other methods, in addition it has an advantage of incremental learning that it learns about the new normal events in an unsupervised manner.