Region of Interest Growing Neural Gas for Real-Time Point Cloud Processing

摘要

This paper proposes a real-time topological structure learning method based on concentrated/distributed sensing for a 2D/3D point cloud. First of all, we explain a modified Growing Neural Gas with Utility (GNG-U2) that can learn the topological structure of 3D space environment and color information simultaneously by using a weight vector. Next, we propose a Region Of Interest Growing Neural Gas (ROl-GNG) for realizing concentrated/distributed sensing in real-time. In ROI-GNG, the discount rates of the accumulated error and utility value are variable according to the situation. We show experimental results of the proposed method and discuss the effectiveness of the proposed method.