Background subtraction is a common method for detecting foreground regions in visual surveillance systems. Since background subtraction is usually the first step of the visual surveillance system, minimizing the processing time and memory consumption is crucial to the performance of entire system. In this paper, we propose two kinds of improved background subtraction methods which consider temporal redundancy and spatial redundancy, respectively. Firstly, we propose a fast cascaded background subtraction method (FCB) which combines frame difference and Gaussian mixture model. By exploiting temporal redundancy in a cascaded fashion, we can reduce the processing time while maintaining the high detection accuracy. Secondly, we propose a memory-efficient cluster-level background subtraction (MCB) to utilize spatial redundancy. By combining similar pixel-level background models into one cluster-level background model, we can considerably reduce the memory consumption for storing the background model. Finally, we combine two methods to take advantages of temporal and spatial redundancy at the same time.