Complex Background Subtraction by Pursuing Dynamic Spatio-Temporal Models

摘要: Although it has been widely discussed in video surveillance, background subtraction is still an open problem the context of complex scenarios, e.g., dynamic backgrounds, illumination variations, and indistinct foreground objects. To address these challenges, we propose effective method by learning maintaining array texture models within spatio-temporal representations. At any location scene, extract a sequence regular bricks, i.e., volumes spanning over both spatial temporal domain. The modeling thus posed as pursuing subspaces bricks while adapting scene variations. For each pursue subspace employing auto regressive moving average model that jointly characterizes appearance consistency coherence observations. During online processing, incrementally update to cope with disturbances from objects changes. In experiments, validate proposed several show superior performances other state-of-the-art approaches subtraction. empirical studies parameter setting component analysis are presented well.