Background modeling is an essential computer vision task in surveillance systems. In background modeling, dealing with non-stationary background is an important problem along with real time performance and memory requirement. Corner-based background modeling is one of the methods that seek to increase the speed, reduce the memory requirement and deals with time varying dynamic backgrounds. In this paper, the corner-based background modeling method is investigated to increase the performance of the algorithm further. The method is simple, computational and memory efficient and effective in modeling dynamic backgrounds. One of the features of this method is its sparse background model. However, when applied to simple scenes that contain large smooth regions, the foreground detection sensitivity decreases and the sparsity of the background model is lost, due to which the memory requirement also increases. Therefore, to eliminate these weaknesses an analysis of the background model is introduced that efficiently detects smooth and textured regions in the scene. The algorithm is then adjusted according to the type of region in order to maintain the model sparsity and increase detection sensitivity. Experimental results show that the proposed method increases the number of correctly detected foreground features by an average 20% for the test sequences we used. The increase in number of foreground features will in turn provide more information to the higher level computer vision modules. The proposed method also maintains the model sparsity over time which in turn results in reduced memory requirement.