Background Initialization Based on Adaptive Online Low-rank Subspace Learning

• Main Authors: Han, Guang, Zhang, Guanghao, Cai, Xi
• Format: Conference Proceeding
• Language: English
• Subjects: Adaptation models; adaptive weighting parameter; background initialization; low-rank subspace learning; Matrix decomposition; Measurement; online robust principal component analysis; Principal component analysis; Sparse matrices; Vehicle dynamics; Video sequences
• ISSN: 2164-5221
• DOI: 10.1109/ICSP48669.2020.9320960

Background initialization is to estimate an appropriate representation for background of a scene, and has a decisive role in determining the performance of background subtraction. Background initialization based on low-rank subspace learning can obtain the background by learning the low-rank subspace. However, most of these methods are batch- based methods requiring heavy memory cost and unable to adapt to dynamic scenes. Accordingly, in this paper, we propose a background initialization method based on adaptive online low-rank subspace learning. The low-rank background subspace is estimated by online robust principal component analysis (PCA) in an online manner. An adaptive weighting parameter is utilized in the online robust PCA to enhance its ability to dynamically model the background. Experimental results demonstrate that, the proposed method can effectively gain the backgrounds of dynamic scenes.