Indoor environments typically consist of sets of connected room-like spaces. We present a local technique that uses range data to detect these spaces during navigation. Our technique includes two parts: segmentation, which isolates room-like spaces and detects when the robot has entered a new one; and feature extraction, which associates each space with a set of geometric features useful for navigation or recognition. Many such features can be considered: here we propose a new method to compute width and length of a rectangular room in a way which is largely invariant with respect to the configuration of the furniture. We report experimental results that show the performance of our technique, and hint at a possible use of this technique for coarse localization on a topological map.