Automatic built-up areas detection from remote sensing images has attracted considerable research interest, due to its crucial roles in various applications. As far as built-up areas detection, the corner density map to predict the presence of the built-up areas has been widely adopted, but the calculation is generally time-consuming. In addition, the density map is just segmented by a statistical threshold, resulting in that the accurate boundaries of the built-up areas are unachievable. In order to address these issues, this paper proposes a novel built-up areas detection approach. Instead of pixel units, our approach takes the superpixel-based image partitions as the primary calculation units, which benefits to improve the computational efficiency and visual organization performance. Based on the superpixel-based units, this paper first proposes a sparse corner voting method for accelerating the production of corner density map. Then, Cauchy graph embedding optimization is presented to cope with the problem of segmenting the density map, which can preserve the well-defined boundaries of built-up areas. A diverse and representative test set including 2.1-m resolution ZY3 imagery, 2.0-m resolution GF1 imagery, 1.0-m resolution IKONOS imagery, and 0.61-m resolution QUICKBIRD imagery is collected. Experimental results on these test images show that our proposed approach is robust to sensor and resolution variation, and can outperform state-of-the-art approaches remarkably.