Using GPUs to accelerate RGB-D mapping

Since the low-cost depth sensors have become widely available, RGB-D mapping is getting more popular in mobile robotics. These sensors provide large amounts of data providing images and depth information. To build 3D maps on-line the high speed data registration alogrithm is needed. To solve this problem we propose the GPGPU implementation of the FAST and BFROST matching algorithms using both RGB image and normal vectors visualization image. We compute the initial transformation usisng RANSAC and HORN closed-form solution algorithm. Finally, we use semantic information to improve the GPGPU ICP and to get an accurate model of the environment.