Dictionary learning for robust background modeling

摘要

Background subtraction is a fundamental task in many computer vision applications, such as robotics and automated surveillance systems. The performance of high-level visions tasks such as object detection and tracking is dependent on effective foreground detection techniques. In this paper, we propose a novel background modeling algorithm that represents the background as a linear combination of dictionary atoms and the foreground as a sparse error, when one uses the respective set of dictionary atoms as basis elements to linearly approximate/reconstruct a new image. The dictionary atoms represent variations of the background model, and are learned from the training frames. The sparse foreground estimation during the training and performance phases is formulated as a Lasso [1] problem, while the dictionary update step in the training phase is motivated from the K-SVD algorithm [2]. Our proposed method works well in the presence of foreground in the training frames, and also gives the foreground masks for the training frames as a by-product of the batch training phase. Experimental validation is provided on standard datasets with ground truth information, and the receiver operating characteristic (ROC) curves are shown.