Online algorithm for foreground detection based on incremental nonnegative matrix factorization
Rong'an Chen, Hui Li · 2016
Foreground detection is widely used in many applications of computer vision and artificial intelligence. In this paper, a novel online algorithm of detecting moving objects in complex scenes is proposed based on incremental nonnegative matrix factorization (INMF). In this algorithm, a new video frame is modeled as a linear combination of basis vectors of background subspace, plus a sparse term which denotes the foreground. In order to take advantage of nonnegative matrix factorization methods, the nonnegative constraints are applied on the update for the coefficient vector and background subspace. Traditionally in many algorithms, the foreground is considered as a pixel-wise sparse term although actually it is structured sparse. So in this paper structured sparsity-inducing norm is introduced to regularize foreground. Besides, in order to efficiently adapt to the dynamically changing background, the contribution of old and new frames can be balanced when updating the background subspace. Experiments on widely used dataset CDnet show that our algorithm has strong adaptability to complex scenes and can detect foreground accurately.