1) The document proposes a background subtraction technique called multi-hypothesis mixture-of-Gaussians (MH-MOG) that generates multiple detection hypotheses from a single background model to achieve superior foreground detection quality for real-time video analytics. 2) MH-MOG maintains a single background pixel model using a perception-aware mixture-of-Gaussians and generates two independent detection hypotheses: a perception-inspired hypothesis and a probabilistic hypothesis. 3) It computes confidence levels for each hypothesis and uses this information in a detection algorithm to produce a high quality foreground mask by maximizing the complementary strengths of the two hypotheses, outperforming conventional techniques that use multiple background models.