Robust abandoned object detection using dual foregrounds

Abstract

As an alternative to the tracking-based approaches that heavilydepend on accurate detection of moving objects, which often failfor crowded scenarios, we present a pixelwise method that employsdual foregrounds to extract temporally static image regions.Depending on the application, these regions indicate objects thatdo not constitute the original background but were brought intothe scene at a subsequent time, such as abandoned and removeditems, illegally parked vehicles. We construct separate long- andshort-term backgrounds that are implemented as pixelwisemultivariate Gaussian models. Background parameters are adaptedonline using a Bayesian update mechanism imposed at differentlearning rates. By comparing each frame with these models, weestimate two foregrounds. We infer an evidence score at each pixelby applying a set of hypotheses on the foreground responses, andthen aggregate the evidence in time to provide temporalconsistency. Unlike optical flow-based approaches that smearboundaries, our method can accurately segment out objects even ifthey are fully occluded. It does not require on-site training tocompensate for particular imaging conditions. While having alow-computational load, it readily lends itself to parallelizationif further speed improvement is necessary.