Abstract
Background modeling and subtraction is a fundamental re- search topic in computer vision. Pixel-level background model uses a Gaussian mixture model (GMM) or kernel density estimation to repre- sent the distribution of each pixel value. Each pixel will be process in- dependently and thus is very efficient. However, it is not robust to noise due to sudden illumination changes. Region-based background model uses local texture information around a pixel to suppress the noise but is vulnerable to periodic changes of pixel values and is relatively slow. A straightforward combination of the two cannot maintain the advan- tages of the two. This paper proposes a real-time integration based on robust estimator. Recent efficient minimum spanning tree based aggre- gation technique is used to enable robust estimators like M-smoother to run in real time and effectively suppress the noisy background estimates obtained from Gaussian mixture models. The refined background esti- mates are then used to update the Gaussian mixture models at each pixel location. Additionally, optical flow estimation can be used to track the foreground pixels and integrated with a temporal M -smoother to ensure temporally-consistent background subtraction. The experimental results are evaluated on both synthetic and real-world benchmarks, showing that our algorithm is the top performer.