Background modelling and subtraction is an essential component in motion analysis with wide range of applications in computer vision, whereas the task becomes more challenging in context of complex scenarios such as dynamic backgrounds. In this paper, we address the problem of modelling dynamic backgrounds in online tensor leaning framework. We use Tucker decomposition to model thespatio-temporal correlation of video background. To facilitate the online execution of foreground detection, we incrementally update the subspace factor matrices and core tensor by using the N-mode SVD. For the upcoming frame, the estimate of new basis matrix is updated, whereas the contents from last observation are removed. Similarity measure based on pixel values is carried out to produce the foreground mask. Visual analysis on video datasets has revealed that the proposed approach is well suited against dynamically varying backgrounds. Our quantitative results show that the proposed strategy is superior to state-of-the-art methods.