Abnormal behavior analysis using LDA

An abnormal behavior analysis method based on bag-of-words model and Latent Dirichlet Allocation was proposed. A video is representedas a collection of spatial-temporal words by using a Gabor filter to extractspace-time interest points. As the noise of the camera and the movement in neighboring frame can also be detected by the filter, we remove the redundant points. Using the 3D-sift features to describe the points and k-means algorithm to cluster the points into words. A video is processing as a text, and the interest points are the words. The LDA assuming that there are some latent topics between the words and the text, and analyzingthe video by the distributing of topic. Although our method uses fewer interest points, it can get a good accuracy. The algorithm was tested on the challenging datasets: Weizmann human action datasets and KTH datasets.