Deep Learning and OcTree-GPU-Based ICP for Efficient 6D Model Registration of Large Objects

In this paper, the shape alignment method that takes into account large objects with very limited visibility and significant occlusions utilizing an octree data structure on the GPU is presented. The proposed algorithm relies on the offline computed 3D model of the object and an initial estimation of it’s pose using a deep learning technique to detect key features of the object, in order to improve the accuracy and the speed of the registration process. The final aligned pose is achieved by computing the iterative closest point algorithm on GPU utilizing octree, starting from the initial estimated pose. To highlight the application of the proposed method, autonomous robotic tasks requiring interaction with washing machine is discussed. Finally, the performance in terms of both speed and accuracy of the different implementations of the algorithm on the CPU and GPU, as well as with and without augmented octree neighbourhood search is provided.