Dense visual simultaneous localisation and mapping (SLAM) systems can produce 3D reconstructions that are digital facsimiles of the physical space they describe. Systems that can produce dense maps with this level of fidelity in real time provide foundational spatial reasoning capabilities for many downstream tasks in autonomous robotics. Over the past 15 years, mapping small scale, indoor environments, such as desks and buildings, with a single slow moving, hand-held sensor has been one of the central focuses of dense visual SLAM research. However, most dense visual SLAM systems exhibit a number of limitations which mean they cannot be directly applied in collaborative or outdoors settings. The contribution of this thesis is to address these limitations with the development of new systems and algorithms for collaborative dense mapping, efficient dense alternation and outdoors operation with fast camera motion and wide field of view (FOV) cameras. We use ElasticFusion, a state-of-the-art dense SLAM system, as our starting point where each of these contributions is implemented as a novel extension to the system. We first present a collaborative dense SLAM system that allows a number of cameras starting with unknown initial relative positions to maintain local maps with the original ElasticFusion algorithm. Visual place recognition across local maps results in constraints that allow maps to be aligned into a common global reference frame, facilitating collaborative mapping and tracking of multiple cameras within a shared map. Within dense alternation based SLAM systems, the standard approach is to fuse every frame into the dense model without considering whether the information contained within the frame is already captured by the dense map and therefore redundant. As the number of cameras or the scale of the map increases, this approach becomes inefficient. In our second contribution, we address this inefficiency by introducing a novel information theoretic approach to keyframe selection that allows the system to avoid processing redundant information. We implement the procedure within ElasticFusion, demonstrating a marked reduction in the number of frames required by the system to estimate an accurate, denoised surface reconstruction. Before dense SLAM techniques can be applied in outdoor scenarios we must first address their reliance on active depth cameras, and their lack of suitability to fast camera motion. In our third contribution we present an outdoor dense SLAM system. The system overcomes the need for an active sensor by employing neural network-based depth inference to predict the geometry of the scene as it appears in each image. To address the issue of camera tracking during fast motion we employ a hybrid architecture, combining elements of both dense and sparse SLAM systems to perform camera tracking and to achieve globally consistent dense mapping. Automotive applications present a particularly important setting for dense visual SLAM systems. Such applications are characterised by their use of wide FOV cameras and are therefore not accurately modelled by the standard pinhole camera model. The fourth contribution of this thesis is to extend the above hybrid sparse-dense monocular SLAM system to cater for large FOV fisheye imagery. This is achieved by reformulating the mapping pipeline in terms of the Kannala-Brandt fisheye camera model. To estimate depth, we introduce a new version of the PackNet depth estimation neural network (Guizilini et al., 2020) adapted for fisheye inputs. To demonstrate the effectiveness of our contributions, we present experimental results, computed by processing the synthetic ICL-NUIM dataset of Handa et al. (2014) as well as the real-world TUM-RGBD dataset of Sturm et al. (2012). For outdoor SLAM we show the results of our system processing the autonomous driving KITTI and KITTI-360 datasets of Geiger et al. (2012a) and Liao et al. (2021) respectively.