Growing Neural Gasに基づく3次元知覚システムにおける体積推定を用いたノード密度調整手法

抄録

With the recent development of robot technology, there are high expectations for autonomous robots. Specifically, the ability to perceive the 3D spatial structure in real-time is essential for autonomous robots. In this research, we propose a method for spatial recognition based on Growing Neural Gas (GNG), which learns the spatial structure of 3D point cloud data, according to the granularity of tasks given to the robots. Our aim is to establish an efficient methodology for perceiving three-dimensional space. In this paper, we estimate the node density by using Principal Component Analysis (PCA) for each cluster and estimating the volume of each cluster from the eigenvalues. We then propose a method to adjust the node density of each cluster based on the node density calculated by the estimated volume and the attenuation rate of the accumulation error, which is a reference for adding nodes in GNG. Finally, we verify the effectiveness of the proposed method through numerical experiments using simulation data and 3D point cloud data measured by an RGB-D camera.