부산대학교 도서관

Robust background subtraction (BS) is essentialfor high quality foreground detection in most video analyticssystems. Recent BS techniques achieve superior detection qualitymostly by exploiting the complementary strengths of multiplebackground models or processing stages. Consequently, these techniques fail to meet the operational requirements ofreal-time video analytics due to high computational overheadwhere BS is just the primary processing task. In this paper, we propose a new BS technique, named multi-hypothesismixture-of-Gaussians (MH-MOG), suitable for real-time videoanalytics. The essential idea is to maintain a single backgroundmodel based on perception-aware mixture-of-Gaussians andthen, generating multiple detection hypotheses with differentprocessing bases. Finally, only during the detection stage, thecomplementary strengths of the hypotheses are exploited toachieve superior detection quality without significant computationaloverhead. Comprehensive experimental evaluationvalidates the efficacy of MH-MOG.