Node Density Adjusting Method of Growing Neural Gas

Abstract

Growing Neural Gas (GNG) based space perception systems have been proposed by many researches since the GNG can learn a geometric structure and generate a topological structure simultaneously. However, these proposed methods assumed the simple perceptual task and cannot apply the multi perception tasks because theses proposed methods use the background extraction. For solving this problem, our previous research proposed Region of Interest GNG (ROI-GNG). ROI-GNG can build concentrated/distributed geometric structure according the degree of attention by controlling the discount rate of the accumulated error according to the degree of attention. However, ROI-GNG has a problem about the stability of the node addition and deletion capabilities. In this paper, therefore, we propose a modified ROI-GNG based on GNG with targeting (GNG-T) that can stably learn the geometric space of the vision sensor by utilizing the d-shortest confidence interval to the algorithm related with the node addition and deletion for improving the stability. In addition, we designed the parameter of the related importance for applying modified ROI-GNG to image data. Finally, we show experimental results of the proposed method and discuss the effectiveness of the proposed method.