Robust and fast mapping based on direct methods

Abstract

Simultaneous localization and mapping technology is the key technology of robot. It's now still challenging due to huge computational cost, rapid motion, non-texture environment, illumination fluctuations, etc. In this paper, we propose a SLAM algorithm based on direct method similar to LSD SLAM. In our method, we adopt constant-velocity motion model to predict the camera pose of next frame and we demonstrate the model's function on avoiding tracking lost. Besides, we adopt inverse compositional method rather than forward additive method used by LSD SLAM to align image and experiments show that ours is more efficient. In addition, we optimize the map on Sim(3) to avoid scale-drift of monocular SLAM.