Robust median background subtraction for embedded vision platforms

Abstract

In this paper two modification to the standard median background subtraction algorithm are proposed, which allows for higher accuracy, real-time performance and entirely integer arithmetic implementation. First, instead of a single pixel intensity, cumulative sums of intensities in patches of various size around each pixel are computed. Thanks to this the method is less sensitive to spurious variations in video and, in result, it is more insensitive to false positives. The second modification is the background detection inference rule, which we propose to base on the modified median absolute variation analysis. This rule further avoids false positives and does not depend on sensitive thresholds. We show that the proposed method has better accuracy than the classical median method, as well as it favorably compares to the group of subspace based background subtraction methods. The proposed method also allows real-time operation on HD video streams.