A Fast and Effective Background Modeling Method Based on Adaptive Clustering and Motion Analysis

Background subtraction is vital for many computer vision tasks such as moving object detection, objects counting and object tracking, while an accurate background modeling can effectively separate foreground objects from background. In this paper, we proposed a fast and effective background modeling method based on adaptive clustering by outliers elimination. For sampled N recent pixels, we eliminate those outlier pixels which have insufficient neighbors within a given distance threshold value. For background updating, we uniformly mix reference pixels (non-outliers) and old background pixel. We calculate Manhattan distance of inter-pixels in RGB color space for fast speed. Moreover, we are inspired by the fact that moving objects (including swaying ones) should not be used for background updating during each single pixel modeling process of whole background. Thus, we present an enhancement technique by ignoring motion pixels contribution to background modeling. Pixel block's motion vector is determined by block matching and the union of all moving pixel blocks is skipped for background updating. Several experiments are done for verifying our algorithm in different scenes, the results demonstrate that our method is efficient and effective and adaptive.