Random Block Background Modeling for Foreground Detection in UHD videos

摘要

Conventional foreground detection methods can take hours to detect objects in a single 4K Ultra High Definition (UHD) frame and their memory requirement is too high to be used without a huge investment in dedicated hardware systems. The proposed Random Block Background Modeling (RBBM) is a spatio-temporal method designed to update quickly the background image of UHD videos. By dividing the image into Mega-Blocks, themselves containing smaller Sub-Blocks and by using small randomly selected Sub-Blocks at each frame through a Gaussian average, the RBBM can accelerate the background modeling. Then, the RBBM is used in combination with a Block Propagative Background Subtraction method to detect the foreground. The proposed RBBM method has been compared with multiple other state-of-the-art works on 4 categories of UHD 4K scenes. The RBBM shows the best quality performances, the best ratio processing time per pixel/quality and a low memory requirement.