This thesis concerns the problem of providing a mobile robot with detailed perception of its local environment using a passive, monocular camera. We embrace the paradigm of dense visual SLAM and bring it to the domain of small, low-cost robots. This enables us to directly use information collected from all pixels in an image and create dense reconstructions of environments. We present a complete and self-contained perception system that allows a mobile robot to estimate its ego-motion, perform infrastructure-free auto-calibration and build, in real-time, a detailed map of its environment in the form of a height map from a single, monocular camera. Our system is capable of providing a robot with accurate information in a form directly suitable for local navigation and obstacle avoidance. By adopting more restrictive, task-oriented models and using the domain knowledge about our applications we were able to improve performance and robustness. Furthermore, when designing our algorithms, we put a great emphasis on methods that can be efficiently and in a straightforward manner implemented on parallel architectures, and therefore we can achieve excellent scalability in terms of resolution of input images and environment representation. We believe that this work offers a promising route to a truly usable real-time monocular dense SLAM system for mobile robots.