Deep learning based 3D object recognition using geometric shapes and sizes

Abstract

For many robotics and VR / AR applications, 3D video is an easy-to-use input source. However, point cloud analysis is a difficult task because it is difficult to capture the shape implied by irregular points. In this paper, a point cloud is created from the image acquired through the RGB-D sensor, and geometric shape classification and size information such as plane, sphere, cylinder, cone, and toroid are extracted through FindSurface. The size information derived from this is input to the basic data set to train in the machine learning module, and the object is recognized. Since 3D object recognition is performed through simple machine learning using only geometric shape and size information, it has a great advantage in terms of time and cost compared to deep learning using voxel or point cloud. Experimental results show that our model is noise resistant and outperforms the latest technology in 3D point cloud classification