Understanding 3D Shapes Being in Motion

This paper concerns a classification problem of 3D shapes being in motion. The goal is to develop the system with real-time capabilities to distinguish basic shapes (corners, planes, cones, spheres etc.) that are moving in front of RGB-D sensor. It is introduced an improvement of SoA algorithms (normal vector computation using PCA Principal Component Analysis and SVD Singular Value Decomposition, PFH – Point Feature Histogram) based on GPGPU (General Purpose Graphic Processor Unit) computation. This approach guarantee on-line computation of normal vectors, unfortunately computation time of the PFH for each normal vector is still a challenge to obtain on-line capabilities, therefore in this paper it is shown how to find a region of movement and to perform the classification process assuming the decreased amount of data. Proposed approach can be a starting point for further developments of the systems able to recognize the objects in the dynamic environments.