Universal Multimode Background Subtraction

Abstract

In this paper, we present a complete change detection system named multimode background subtraction. The universal nature of system allows it to robustly handle multitude of challenges associated with video change detection, such as illumination changes, dynamic background, camera jitter, and moving camera. The system comprises multiple innovative mechanisms in background modeling, model update, pixel classification, and the use of multiple color spaces. The system first creates multiple background models of the scene followed by an initial foreground/background probability estimation for each pixel. Next, the image pixels are merged together to form megapixels, which are used to spatially denoise the initial probability estimates to generate binary masks for both RGB and YCbCr color spaces. The masks generated after processing these input images are then combined to separate foreground pixels from the background.

Existing System

Split Gaussian Model, Bayesian Modeling.

Proposed System

In this paper, we propose a BS system that is robust against various challenges associated with real world videos. The proposed approach uses a Background Model Bank (BMB) that comprises of multiple Background (BG) models of the scene. To separate foreground pixels from changing background pixels caused by scene variations or camera itself, we apply Mega-Pixel (MP) based spatial denoising to pixel level probability estimates on different color spaces to obtain multiple Foreground (FG) masks. They are then combined to produce a final output FG mask. The major contribution of this paper is a universal background subtraction system called Multimode Background Subtraction (MBS) with following major innovations: Background Model Bank (BMB), model update mechanism, MP-based spatial denoising of pixel-based probability estimates, fusion of multiple binary masks, and use of multiple color spaces for BS process. We have presented a universal BG subtraction system that exploits multiple BG models and computationally inexpensive pixel-level comparison to generate initial probability estimates, which undergo spatial denoising by forming MPs. To separate vision tasks based on illumination conditions, we use RGB and Y color channels to for low light vision and CbCr for bright light to provide more accurate foreground segmentation. The introduction of FG dependent model update mechanism eliminates the need to tune parameters for every test sequence.

Architecture

Binary classification and mask generation

Universal Multimode Subtraction System