Statistical spatial multi-pixel-pair model for object detection

A novel robust model for background subtraction under complex scenes is proposed. Unlike the previous works, it utilizes multiple pixel-pairs which exhibit a stable co-occurrence relation. In training progress, the support pixels are screened by utilizing temporal covariance matrix, and the spatial distributions of support pixels are optimized by spatial sampling based on K-means clustering, in order to balance high co-occurrence and relaxed spatial distribution. Then in detection progress, with a parametrized condition, the background model performs robust and accurate detections, under two challenging datasets (PETS-2001 and AIST-INDOOR).