Camera tracking for SLAM in deformable maps

Abstract

The current SLAM algorithms cannot work without assuming rigidity. We propose the first real-time tracking thread for monocular VSLAM systems that manages deformable scenes. It is based on top of the Shape-from-Template (SfT) methods to code the scene deformation model. Our proposal is a sequential method that manages efficiently large templates, i.e. deformable maps estimating at the same time the camera pose and deformation. It also can be relocated in case of tracking loss. We have created a new dataset to evaluate our system. Our results show the robustness of the method in deformable environments while running in real time with errors under 3% in depth estimation.