Object recognition in noisy RGB-D data

Abstract

The object recognition task on 3D scenes is a growing research field that faces some problems relative to the use of 3D point clouds. In this work, we focus on dealing with noisy clouds through the use of the Growing Neural Gas (GNG) network filtering algorithm. Another challenge is the selection of the right keypoints detection method, that allows to identify a model into a scene cloud. The GNG method is able to represent the input data with a desired resolution while preserving the topology of the input space. Experiments show how the introduction of the GNG method yields better recognitions results than others filtering algorithms when noise is present.