Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
谷藤 祐太; 川端 邦明
Proceedings of International Workshop on Nonlinear Circuits, Communications and Signal Processing (NCSP 2020) (Internet), p.246 - 249, 2020/02
In this paper, we describe a method of acquiring learning data 3D point cloud data as learning data for deep learning using a simulator. Generally, a lot of data is necessary for building classifiers by deep learning approach. By using a simulator, various measurement conditions can be set thus, it is expected to collect variety of data for building high performance classifier. Data collection was conducted by virtual measurement using a mobile robot model and a sensor model. As a feasibility study of evaluating classification performance, we performed a simple identification experiment to confirm performance and applicability to actual measurement data. As a result, a high identification rate of 89 percent to three categories was obtained.
谷藤 祐太; 川端 邦明; 羽成 敏秀
Proceedings of 2019 IEEE Region Ten Conference (TENCON 2019) (Internet), p.36 - 40, 2019/10
This paper describes a Graphical User Interface (GUI) based operation system for building a classifier based on deep learning and verifying its categorization performance. Currently, we build a structure discrimination method based on deep learning with 3D point cloud to support status awareness of the operator of remotely controlled robot. For building a powerful classifier, the operations like "collection of learning data", "construction of architecture" and "creation of learning model "are done by trial and error. Therefore, we consider to develop a system to make such complicated operations easier and more efficiently. In this paper, we describe about required functions for helping such operations and explain developed a prototype system in detail.
谷藤 祐太; 川端 邦明
Proceedings of International Topical Workshop on Fukushima Decommissioning Research (FDR 2019) (Internet), 4 Pages, 2019/05
This paper describes about the development of an environment recognition method with point cloud data collected in a dark place like Fukushima Daiichi Nuclear Power Station (FDNPS). We reported the results of a feasibility study of the structure discriminations from LiDAR 3D point cloud data by a deep learning approach. Proposed method utilizes the image data of projected 3D point cloud as input for the classifier instead of coordinate data of 3D points directly. This idea realized to make shorten the learning time without large capacity of the memory for the computations. We selected five kinds of structures (Stairs, Pipe, Grating, Switchboard and Valve) commonly appeared in the general plant as a discrimination subjects for evaluating proposed method. As a result, the classifier showed an accuracy of 99.6% to five categories and we could confirm the validity of proposed method for the structure discrimination.