Incremental and adaptive front-end fusion

This paper describes an incremental and adaptive front-end fusion system capable of providing accurate 3D models of the world using a single RGB camera. This is ideal for robotic platforms operating in indoor environments, where there is a need for high fidelity in the reconstruction of the world but at a low memory usage and computational cost. Our algorithm expands on previous volumetric variational approaches for 3D reconstruction by providing two main key features. The first is a novel incremental method for updating the cost volume which removes the need of keeping hundreds of multi-view comparison images, thus reducing the overall processing time and memory storage of the system. The second feature is a method for dynamically adapting the minimum and maximum depth limits of the cost volume as it adjusts to changes in scene depth, thus achieving optimum resolution in the 3D reconstruction.