Short term-Long term dual model Background Subtractor
Arun Varghese, G. Sreelekha · 2015
Background subtraction is the most common technique to segment moving objects in a video sequence. We propose a data driven background subtraction method which uses two background models for each pixel: a long term model and a short term model. The long term model captures the long term evolution of the background while the short term model adapts quickly to rapidly changing background conditions like swaying tree leaves or camera jitter. Each model comprises of a collection of previously observed pixel values. Two segmentation maps are generated based on whether the current pixel value finds the required number of matches with the samples in the corresponding model. The final segmentation mask is obtained as an intersection of the two. Evaluation tests on the public CDnet dataset shows improved performance compared to popular methods.