Fast Feature-Less Quaternion-based Particle Swarm Optimization for Object Pose Estimation From RGB-D Images

We present a novel quaternion-based formulation of Particle Swarm Optimization for pose estimation which, differently from other approaches, does not rely on image features or machine learning. The quaternion formulation avoids the gimbal lock problem, and the objective function is based on raw 2D depth information only, under the assumption that the object region is segmented from the background. This makes the algorithm suitable for pose estimation of objects with large variety in appearance, from lack of texture to strong textures, for the task of robotic grasping. We find candidate object regions using a graph-based image segmentation approach that integrates color and depth information, but the PSO is agnostic to the segmentation algorithm used. The algorithm is implemented on GPU, and the nature of the objective function allows high parallelization. We test the approach on different publicly available RGB-D object datasets, discuss the results and compare them with other existing methods.