Archive | 2019
Background Subtraction by Difference Clustering
Abstract
Previous approaches to background subtraction typically considered the problem as a classification of pixels over time. We frame the problem as clustering the difference vectors between pixels in the current frame and in the background image set, and present a novel background subtraction method called Difference Clustering. This not only saves computational time, but also achieves high Pr and Fm values for accuracy. In particular, the difference between the current frame and the background image set is extracted using the quartile method for clustering. Compared to traditional k-means model to generate k clusters, our quartile method needs only 2 clusters. Moreover, traditional k-means clustering models need to update the means until convergence, which is time-consuming. In contrast, our quartile method finds the final means directly to reduce the numbers of iterations and computational time, resulting in a real-time algorithm. Experiments on several videos from standard benchmarks demonstrate that our proposed approach achieves promising results compared to several previous methods.