Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
Initialising ...
羽成 敏秀; 中村 啓太*; 今渕 貴志; 川端 邦明
計測と制御(インターネット), 65(6), p.438 - 443, 2026/06
廃炉作業が進められている福島第一原子力発電所では高い放射線環境で様々な調査や作業を行うため、様々な遠隔技術が活用されている。廃炉作業環境下で用いられている遠隔技術においてオペレータの空間認知の不良性が問題視されており、それを改善する手法として作業環境の立体モデルの提示が有望視されている。本稿では、画像に基づいた立体モデリングおよびシミュレーション技術の活用やAIを利用した立体モデリング技術について紹介するとともに、廃炉作業への適用に関する課題についても述べる。
谷藤 祐太; 今渕 貴志; 川端 邦明
AI・データサイエンス論文集(インターネット), 7(1), p.64 - 75, 2026/05
本研究では、深層学習を用いたノイズ除去および画像鮮明化を現場作業に適用可能とするため、複数のモデルをGraphical User Interface (GUI)上で柔軟に切り替えられる画像復元システムを開発した。提案システムは、Noise2Noiseを中心に、SRCNN、FSRCNN、RED-Netの4種類のネットワークを統合しており、ユーザは作業環境に応じて最適な処理を選択できる。評価実験では、福島第一原子力発電所で取得された低品質画像に対し、Noise2Noiseが特に優れたノイズ低減性能を示し、画質を向上させることを確認した。また、GUI設計により従来のCharacter User Interface (CUI)ベース操作と比較して操作手順を50%、処理時間を40%削減でき、作業効率の大幅な改善を実現した。さらに、各手法の比較実験により、提案システムが実用性と柔軟性を兼ね備えていることを示した。
高橋 弘毅*; 加藤 徹*; 山下 圏*; 土井 章男*; 今渕 貴志
Artificial Life and Robotics, 31(2), p.598 - 610, 2026/05
To create accurate 3D models from video footage, it is essential to use high-quality videos without floating objects that could interfere with the process. In this study, we applied the Segment Anything Model (SAM) and Generative Image Inpainting to enhance the quality of video frames by detecting and removing floating objects on a frame-by-frame basis. The results demonstrated the effectiveness of this approach in detecting and eliminating such objects, contributing to the improvement of video quality.
今渕 貴志; 川端 邦明
Proceedings of 2026 IEEE/SICE International Symposium on System Integration (SII2026) (Internet), p.1100 - 1104, 2026/01
This paper describes a fully automated method for generating 3D pipe models with wall thickness from 3D point cloud data. In decommissioning of Fukushima Daiichi Nuclear Power Station, rapid 3D modeling of plant structures is essential for dose assessment and remote operation planning. In our method, pipe regions are first discriminated, then scanned along three orthogonal axes to extract pipe instances, followed by geometric fitting with wall thickness for use in shielding calculations. Finally, generated models are exported in Computer-Aided Design (CAD) and Building Information Modeling (BIM) formats for using facility management. Our method was validated on a 3D point cloud measured in a mock-up plant environment. We confirmed a high conversion rate and reduced processing time by restricting computation to pipe regions. The outputs models imported into existing software without errors.
谷藤 祐太; 今渕 貴志; 川端 邦明
Proceedings of the 31st International Symposium on Artificial Life and Robotics (AROB 31st 2026), p.1011 - 1016, 2026/01
福島第一原子力発電所(1F)の廃炉作業では、濁りや浮遊物、レンズの汚染、放射線起因ノイズなどにより劣化したロボットカメラ映像を用いて作業を行う必要があり、クリーンな参照画像は得られない。本研究では、Noise2Noise(N2N)によるノイズ除去とFSRCNNによる超解像を組み合わせた軽量な画像復元パイプラインを検討した。その結果、N2N→FSRCNNの構成が、NIQEおよびPIQEによる画質評価指標および主観評価の両面において、ノイズや不自然なテクスチャを抑制しつつ視認性を安定的に向上させることを確認した。
中島 慎介*; Wu, J.*; 羽成 敏秀; 今渕 貴志; 松日楽 信人*; 川端 邦明; An, Q.*; 山下 敦*
Proceedings of IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR2025), p.127 - 132, 2025/10
被引用回数：0 パーセンタイル：0.00(Automation & Control Systems)Fuel debris retrieval is the indispensable task for decommissioning Fukushima Daiichi Nuclear Power Station (FDNPS). One of the significant barriers is the high radiation which deteriorates the operation efficiency. The radiation makes the robot sensors and the electronics get out of order easily. For the acceleration of decommissioning operation, it is required to develop the DX (Digital Transformation) platform to evaluate the sensor performance in the radiation environment. This research presents the approach exploiting a photorealistic game engine. The developed DX platform contributes to evaluate the sensor degradation and image processing algorithm inside the PCV (Primary Container Vessel) of FDNPS.
加藤 徹*; 高橋 弘毅*; 山下 圏*; 土井 章男*; 今渕 貴志
Artificial Life and Robotics, 30(1), p.126 - 135, 2025/02
We have developed a point cloud processing system within the Unreal Engine to analyze changes in large time-series point cloud data collected by laser scanners and extract structured information. Currently, human interaction is required to create CAD data associating with the time-series point cloud data. The Unreal Engine, known for its 3D visualization capabilities, was chosen due to its suitability for data visualization and automation. Our system features a user interface that automates update procedures with a single button press, allowing for efficient evaluation of the interface's effectiveness. The system effectively visualizes structural changes by extracting differences between pre- and post-change data, recognizing shape variations, and meshing the data. Difference extraction involves isolating only the added or deleted point clouds between the two datasets using the K-D tree method. Subsequent shape recognition utilizes pre-prepared training data associated with pipes and tanks, improving accuracy through classification into nine types and leveraging PointNet++ for deep learning recognition. Meshing of the shape-recognized point clouds, particularly those to be added, employs the Ball Pivoting Algorithm (BPA), which was proven effective. Finally, the updated structural data is visualized by color-coding added and removed data in red and blue, respectively, within the Unreal Engine. Despite increased processing time with a higher number of point cloud data, down sampling prior to difference extraction signific reduces the automatic update time, enhancing overall efficiency.
今渕 貴志; 川端 邦明
Artificial Life and Robotics, 30(1), p.184 - 195, 2025/02
This paper describes a 3D point cloud segmentation pipeline that contributes to the efficiency of decommissioning works at the Fukushima Daiichi Nuclear Power Station. For decommissioning works, simulations and calculations for preliminary work planning using 3D structural models are crucial from a safety and efficiency viewpoint. However, 3D modeling works typically require high costs. Therefore, we aim to improve the efficiency of 3D modeling by segmenting geometric shape regions into categories in a 3D point cloud state using deep learning. Our pipeline uses 3D computer-aided design semantics to create a training dataset that reduces annotation costs and helps learn human knowledge. Performance evaluation results show that the discriminator can discriminate major structural categories with high accuracy using deep learning models. However, we confirm that even the state-of-the-art model has limitations in discriminating structures containing similar shapes between categories and structures in categories with a small number of training data. In the analysis of evaluation results, we discuss challenges encountered by our pipeline for practical applications.
今渕 貴志; 羽成 敏秀; 川端 邦明
Proceedings of 2025 IEEE/SICE International Symposium on System Integration (SII2025), p.1416 - 1421, 2025/01
This paper describes a 3D reconstruction based on grouping similar structures for the aim of generating 3D information for understanding the workspace from the images acquired inside the Primary Containment Vessel (PCV) of the Fukushima Daiichi Nuclear Power Station. In the decommissioning works, preliminary surveys are carried out in the PCV, and the workers need to understand the workspace from a large amount of camera images, which requires a great deal of effort. We are currently working on 3D reconstruction from camera images of the PCV, however, one of the challenges is to improve the visibility of reconstructed model containing noise and artifact. In this study, we propose a method of grouping similar structures on the image and utilizing predicted group labels for 3D reconstruction process to highlight structures shapes and to refine 3D modeling. Our key idea is to perform unsupervised segmentation for grouping similar structures that are suitable for images acquired in the PCV because they are difficult to assign correct semantics for unclear structures and the few learning resources. We show on the reasonable performance of our method by validating it by video images of a typical plant environment and survey videos of the PCV taken under adverse conditions such as radiation noise.
羽成 敏秀; 中村 啓太*; 今渕 貴志; 川端 邦明
Journal of Robotics and Mechatronics, 36(6), p.1537 - 1549, 2024/12
本稿では、時系列画像から3次元モデルを効率的に生成するための画像選択法を導入した立体復元処理について述べる。効率的な立体復元に適した画像を得るために、時系列画像中の冗長画像を除去する画像選択法の適用を試みた。提案手法では、オプティカルフローと固定しきい値に基づいて、時系列画像から適切な画像を選択する。その結果、提案手法により、カメラで取得した時系列画像に基づく立体復元処理の計算量を削減することができた。その結果、立体復元精度を一定に保ったまま、立体復元処理の計算量を削減できることを確認した。
中村 啓太*; 馬場 啓多*; 渡部 有隆*; 羽成 敏秀; 松本 拓*; 今渕 貴志; 川端 邦明
Artificial Life and Robotics, 29(4), p.546 - 556, 2024/09
This paper describes a method for integrating multiple dense point clouds using a shared landmark to generate a single real-scale integrated result for photogrammetry. It is difficult to integrate high-density point clouds reconstructed by photogrammetry because the scale differs with each photogrammetry. To solve this problem, this study places a QR code of known sizes, which is a shared landmark, in the reconstruction target environment and divides the reconstruction target environment based on the position of the QR code that is placed. Then, photogrammetry is performed for each divided environment to obtain each high-density point cloud. Finally, we propose a method of scaling each high-density point cloud based on the size of the QR code and aligning each high-density point cloud as a single high-point cloud by partial-to-partial registration. To verify the effectiveness of the method, this paper compares the results obtained by applying all images to photogrammetry with those obtained by the proposed method in terms of accuracy and computation time. In this verification, ideal images generated by simulation and images obtained in real environments are applied to photogrammetry. We clarify the relationship between the number of divided environments, the accuracy of the reconstruction result, and the computation time required for the reconstruction.
川端 邦明; 今渕 貴志; 白崎 令人*; 鈴木 壮一郎; 伊藤 倫太郎; 青木 勇斗; 大森 崇純
ROBOMECH Journal (Internet), 11, p.11_1 - 11_11, 2024/09
This paper describes a measuring unit to realize the synchronous collection of air dose rate and measurement position for efficient dosimetry survey and data logging in a working space. The developed prototype comprises a three-dimensional light detection and ranging-based mapping part and a dosimetry part, which are integrated into a single measurement unit through an embedded computer that installs a ROS (Robot Operating System) framework. The unit can function as a standalone system with embedded batteries. Since it is portable, on-line data gathering in the workspace can be realized, thereby maintaining consistency between the air dose rate and the measurement position. In this paper, we describe the prototype system configuration and the experimental results obtained in the mockup test space and nuclear facility to discuss its performance.
中村 啓太*; 羽成 敏秀; 今渕 貴志; 川端 邦明
Proceedings of 2024 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM 2024), p.7 - 8, 2024/07
写真測量法は、対象物を撮影した複数の画像から3次元的に復元する技術である。実際の撮影では、対象物を撮影するカメラの画角に振動が発生し、写真測量に適した画像を撮影できず、対象物を復元できない場合がある。そこで、この振動をシミュレーション上で乱数を用いて実装し、振動の大きさが写真測量法で得られる復元結果に与える影響を検証した。検証の結果、振動の大きさと3次元復元の成功率との関係が示された。
今渕 貴志; 川端 邦明
Journal of Robotics and Mechatronics, 36(1), p.63 - 70, 2024/02
In the decommissioning of Fukushima Daiichi Nuclear Power Station, radiation dose calculations using a 3D model of the workspace are performed to determine appropriate measures to reduce exposure. However, constructing a 3D model from 3D point cloud is costly. In order to separate the geometrical shape regions on 3D point cloud, we have been developing the structure discrimination method by 3D and 2D deep learning for contributing to 3D modeling automation technology. In this paper, we describe a method for transferring and fusing labels to handle 2D prediction label in 3D space. We propose an exhaustive label fusion method for plant facilities with intricate structures. In evaluation, we applied the method to a mock-up plant dataset and confirmed that it works effectively.
今渕 貴志; 川端 邦明
Proceedings of 2024 IEEE/SICE International Symposium on System Integration (SII2024) (Internet), p.141 - 146, 2024/01
This paper describes a method for volumetric-based semantic 3D modeling from 3D point cloud obtained in a plant environment. In order to calculate a radiation dose distribution of the workspace in decommissioning, the shape, arrangement, materials, and thicknesses of structures are essentially required in addition to dose values. However, it is costly to create such enriched 3D models from 3D point cloud. In this study, we propose a method to create 3D models with structural category and material thickness by combining 2D image-based deep learning and volumetric reconstruction method. To discriminate structures, structural category labels are predicted by a pre-trained 2D semantic segmentation network on projected image created from 3D point cloud. Then, a triangular mesh is generated from the integrated Truncated Signed Distance Function (TSDF) according to prediction labels. In addition, we optimize the TSDF thickness assignment function to reduce surface distance error. Our evaluation reports thickness and surface distance errors when generating meshes with three different structural categories in a mock-up plant environment.
加藤 徹*; 高橋 弘毅*; 山下 圏*; 土井 章男*; 今渕 貴志
Proceedings of 29th International Symposium on Artificial Life and Robotics (AROB 29th 2024) (Internet), p.1097 - 1100, 2024/01
We have researched and developed a point cloud processing system on the Unreal Engine that recognizes changes between large time-series point cloud data measured by a laser scanner and performs structured data extraction. When associating time-series point cloud data with structural information (pipes, tanks, etc.) of each point cloud, CAD data (structural data) is currently created interactively by humans. Unreal Engine is a game engine that excels in visualization of 3D information and is suitable for checking updated data and automating procedures. We developed a user interface that automatically performs a series of update procedures at the touch of a button, and evaluated the effectiveness of the interface.
高橋 弘毅*; 加藤 徹*; 山下 圏*; 土井 章男*; 今渕 貴志
Proceedings of 29th International Symposium on Artificial Life and Robotics (AROB 29th 2024) (Internet), p.1093 - 1096, 2024/01
When performing shape recognition using machine learning from point cloud data, data expansion can be expected to improve accuracy. In this research, we developed a virtual 3D scanner, VCC (Virtual Cloud Creator), that automatically generates data with a distribution close to actual measurement from large-scale point clouds. VCC can generate labeled point cloud data with an appropriate point cloud density from any viewpoint using 3D measured point cloud data and CAD data created from the point cloud data. It was confirmed that shape recognition accuracy was improved by using the augmented data generated by VCC for learning.
今渕 貴志; 川端 邦明
Proceedings of 20th International Conference on Ubiquitous Robots (UR 2023), p.632 - 633, 2023/06
This paper describes a feasibility study of a deep learning-based structure discrimination using data sources measured by different sensor from the sensor used when training the network for automatic creation of 3D spatial information in working environment for decommissioning. In previous work, we have developed a method for estimating geometrical shape regions and category information of structures from high-density 3D point cloud data measured in nuclear facilities. In this paper, we report the discrimination accuracy on low-density 3D point cloud data measured in a testing plant environment using the discriminator trained on high-density and high-precision 3D point cloud data measured in the same space. In addition to this, we verify the improvement of accuracy by removing noise on the surface of the structure caused by the sensor's measurement error range.
今渕 貴志; 川端 邦明
Proceedings of 2023 IEEE/SICE International Symposium on System Integration (SII 2023) (Internet), p.396 - 400, 2023/01
被引用回数：2 パーセンタイル：59.45(Computer Science, Interdisciplinary Applications)This paper describes a method for discrimination of structures in a plant facility by deep learning based on projected images created from a colored 3D point cloud data using a virtual camera system. In order to promote safe and secure decommissioning works, it is important to recognize a radiation condition in the workspace via calculation based on the measured sensor data. In our previous work, we proposed a structure discrimination method by 3D semantic segmentation network to obtain clues required for radiation dose simulation: shape regions for creating 3D shape model and structural category labels for assigning material information. However, in the evaluation, we confirmed that the network trained based on point geometric patterns had limited discrimination performance. In this paper, we introduce deep learning based on projected images created from colored 3D point cloud data to improve the accuracy of our structure discrimination method. The projected images are created by a virtual camera system, and after discrimination, predicted pixel-wise labels are back-projected into the 3D point cloud. In evaluation, we reported the discrimination performance and 2D-3D back projection result of our proposed method.
今渕 貴志; 谷藤 祐太; 川端 邦明
Proceedings of 2022 IEEE/SICE International Symposium on System Integration (SII 2022) (Internet), p.1036 - 1040, 2022/01
被引用回数：3 パーセンタイル：67.06(Computer Science, Interdisciplinary Applications)本稿は、3次元点群データに基づいた深層学習による原子力プラント内構造物の識別手法について報告するものである。安全かつ着実な廃炉作業の推進のためには、取得した作業空間情報に基づいて状態の推定や想定を行うことが重要である。特に空間内の放射線量に関わる情報をシミュレーション等の計算で正確に推定することができれば作業計画の立案等が容易になる。正確な放射線量に関する計算を行うためには、作業空間内に存在する物体の空間形状と材質のデータが必要不可欠であるが、その推定は容易ではない。そこで我々は、原子力プラント内の構造物の種別から材質を推定しうることに着目し、深層学習による構造物識別手法の研究開発に取り組んでいる。本稿では、学習データとして既存プラントの3次元点群データに基づき作成されたCADデータの属性ラベルをボクセルグリッドサンプリングによって当該点群データへ付与する手法を提案し、Hold-out学習による学習精度を示すことでその有用性についての考察を述べる。
