187037 research outputs found
Sort by
궤도상 형상 복원 및 상대항법을 위한 라이다의 강도 측정 정보를 이용한 포인트 클라우드 등록 알고리즘
No full text학위논문(석사) - 한국과학기술원 : 항공우주공학과, 2025.2,[iv, 46 p. :]In a rendezvous and proximity operation (RPO) scenario to an unknown and uncooperative target, simultaneously reconstructing the target's geometry and performing relative navigation of the chaser presents a significant challenge. Some Simultaneous Localization and Mapping (SLAM) algorithms address this problem efficiently by using point cloud registration algorithms such as Iterative Closest Point (ICP). However, ICP is highly sensitive to its initial alignment conditions, where even small misalignment in the initial conditions can result in matching failure. This paper introduces a BRDF ICP, a novel feature-aided ICP for RPO scenarios, which mitigates this sensitivity issue by utilizing point clouds and intensity measurements of light detection and ranging (LiDAR) sensors. To model the behavior of LiDAR intensity measurements in space, an accurate LiDAR Bidirectional Reflectance Distribution Function (BRDF) that considers realistic light characteristics is proposed. Using this LiDAR BRDF model, the BRDF segmentation algorithm is developed to segment the point cloud into distinct materials and components. Additionally, a coarse-to-fine registration framework is designed that provides a reliable initial guess using correspondences between BRDF-segmented features. To evaluate the proposed framework, a simulator capable of generating two channels of LiDAR measurements was built and utilized. The accuracy of the BRDF segmentation algorithm was tested using arbitrary and real parameters by analyzing mean Intersection over Union (mIoU) values. To apply BRDF ICP to the RPO scenario, a simple pose graph optimization (PGO) SLAM approach was utilized for geometry reconstruction and relative navigation of a 2U CubeSat target. The performance of the proposed registration algorithm was further validated through Monte Carlo simulations by comparing the errors in the reconstructed target model and relative pose with those of other ICP algorithms.한국과학기술원 :항공우주공학과
온보드 실시간 컨벡스 최적화를 위한 커스텀 SOCP 솔버 개발
No full text학위논문(석사) - 한국과학기술원 : 항공우주공학과, 2025.2,[iv, 75 p. :]This thesis addresses the development of a customized SOCP solver for onboard real-time convex optimization. In recent years, computational guidance and control(G&C) techniques have emerged to overcome the limitations of traditional methods, that heavily rely on analytic solutions, and to leverage the advanced computational power of modern computers. In particular, convex optimization-based G&C techniques have garnered attention due to their ability to handle various forms of constraints, and guaranteed convergence and global optimality. To realize these techniques, a customized solver capable of deriving real-time solutions on flight computers is essential. In this thesis, we developed a customized solver that can be easily deployed onto onboard computers, and is well-suited for real-time solution derivation. The solver utilizes homogeneous embedding that can directly handle quadratic cost functions, enabling it to address all subsets of SOCP without requiring problem reformulation. Thus, more efficient algorithm can be constructed for problems that involve quadratic cost function. We apply primal-dual interior-point methods(IPM) with Mehrotra's predictor-corrector and Nesterov-Todd scaling. This structure of IPM is employed in most modern IPM-based solvers, demonstrating its performance and robustness. Finally we conducted benchmarking on various problems and compared the results with those of existing solvers. We also conducted numerical experiments on embedded systems to assess the feasibility of the developed solver for convex optimization-based G&C techniques.한국과학기술원 :항공우주공학과
희소 성질 오토인코더를 활용한 컨볼루션 신경망 내의 채널 개념 특징 시각화
No full text학위논문(석사) - 한국과학기술원 : 김재철AI대학원, 2025.2,[iv, 28 p. :]Recently, The number of applications and efficiency of Deep Neural Networks(DNN) have been increased rapidly. The explanations and interpretations on activated features internally of the DNN models are required in the field of research also in society. Following this trends, Researches on explaining Transformer-based Language Models exploit Sparse AutoEncoders(SAE) for linear decomposition of model's features. However, in the field of explaining recent image processing models such as Vision Transformers(ViT) and ConvNeXt based models, which process image inputs as Transformers Language models did, the researches cluster or classify Concept Activation Vectors(CAV) into interpretable structures. But, these methods are less intuitive than feature visualizations. Also, in the researches on the feature visualization, they are highly dependent on the specific input sample image. For addressing this problem, this paper suggests CAV optimization visualization from random noise input into interpretable feature image by optimizing objective function of cosine-similarity between CAV and noise input activations following the definition of CAV. Also, this paper shows visualized features are polysemantical, and suggests a method for decomposition of features by acquiring concept vectors exploiting SAE.한국과학기술원 :김재철AI대학원
4D 의료 영상의 데이터 효율적 보간: 비지도 학습에서 지도 학습 강화까지
No full text학위논문(석사) - 한국과학기술원 : 김재철AI대학원, 2025.2,[iv, 25 p. :]4D medical imaging, which is a series of 3D images with temporal information, plays a pivotal role in clinical settings. However, acquiring 4D medical data presents significant challenges, which stem primarily from concerns such as radiation exposure and lengthy imaging times. Under these constraints, not only is data acquisition itself complex, but increasing the temporal frame rate for individual datasets also becomes difficult. To address these issues, we propose UVI-Net, a novel yet straightforward Unsupervised Volumetric Interpolation network that enables temporal interpolation without relying on intermediate frames, setting it apart from most existing unsupervised methods. Notably, despite being designed for fully unsupervised settings, this approach can also be extended to serve as auxiliary supervision in fully supervised interpolation tasks. Experimental results on benchmark datasets demonstrate UVI-Net's state-of-the-art performance in an unsupervised setting, with further improvements observed when combined with supervised training. Moreover, UVI-Net achieves this enhanced performance with only a single training sample, underscoring its robustness and efficiency in situations with limited supervision. This makes UVI-Net a promising solution for 4D medical interpolation—particularly in situations of data scarcity—that effectively works in both unsupervised and supervised settings.한국과학기술원 :김재철AI대학원
Learning dynamic pick-and-place for quadrupedal mobile manipulator
No full text학위논문(석사) - 한국과학기술원 : 김재철AI대학원, 2025.2,[iii, 32 p. :]본 연구에서는 4족 로봇에 6자유도 로봇팔을 장착한 이동 조작기를 활용하여, 물체의 픽-앤-플레이스 작업을 수행하는 동적인 전신 정책 학습 방법론을 제안한다. 계층적 강화학습 기법을 적용하여 학습자의 행동공간을 효율적으로 구성하였으며, 로봇의 고유 감각 정보와 물체 인식 정보를 시계열 데이터로 통합하여,물체의 질량을 잠재 공간에서 내재적으로 추정할 수 있도록 하였다. 이를 통해 팔의 가반하중을 초과하는3kg 내외의 물체를 안정적으로 조작할 수 있으며, 시뮬레이션 결과 약 5cm의 허용 오차 내에서 학습 범위의조작상황에 대해 85.78%의 픽-앤-플레이스 작업 성공률을 달성하였다. 본 연구는 동일한 플랫폼을 사용하는기존 연구들이 비교적 가벼운 물체의 파지에만 국한된 점을 넘어, 물체의 정확한 배치까지 수행할 수 있는이동 조작 시스템을 구현함으로써 효율적인 픽앤플레이스 작업이 가능한 프레임워크를 제안한다.한국과학기술원 :김재철AI대학원
거리 변화에 둔감한 무선 전력 및 데이터 송수신 시스템
No full text학위논문(석사) - 한국과학기술원 : 바이오및뇌공학과, 2025.2,[iv, 41 p. :]This study proposes a system capable of seamless uplink data communication over a single wireless power and data transmission link (WPDT), regardless of distance variations, while incorporating dual-output rectification technology using a shared inductor. To ensure stable operation despite changes in the external (Tx)-internal (Rx) distance, the system utilizes the frequency splitting region, where more energy is received as the distance increases, and employs a dummy capacitor to enable stable transmission and reception of high data rates. The amplitude of the current flowing through the Tx LC tank varies with the distance of the link, and the system changes the current direction of the Rx LC tank accordingly to achieve stable data transmission and insensitivity to distance changes. To address these issues, active current direction switching and dummy capacitors are used to minimize oscillations at short distances and resolve the low modulation index at longer distances, maximizing the high-speed uplink data rate. This study demonstrates a system capable of delivering power to two independent loads via a single link at distances ranging from 10 mm to 42 mm while maintaining a consistent uplink data rate of 3.87 Mbps across all distances.한국과학기술원 :바이오및뇌공학과
다중 개체 확산 강조 영상 정보를 이용한 그래프 합성곱 신경망 기반 파킨슨병 운동 증상 진행 예측
No full text학위논문(석사) - 한국과학기술원 : 바이오및뇌공학과, 2025.2,[iv, 68 p. :]Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor impairments linked
to disruptions in structural brain connectivity. Accurate prediction of motor symptom progression, as quantified
by the Unified Parkinson’s Disease Rating Scale Part III (UPDRS-III), is critical for personalized treatment planning and early intervention. In this study, we propose a novel hierarchical graph convolutional network (GCN) framework that uniquely combines subject-level GCNs (s-GCNs) to model region-to-region connectivity, a long short-term memory (LSTM) network for capturing temporal trends, and a population-level GCN (p-GCN) to incorporate inter-subject relationships. This integration not only surpasses SOTA models in predictive accuracy but also, through GNNExplainer, provides interpretable insights into motor symptom progression by identifying biologically meaningful regions of interest (ROIs). Our framework comprises three key components. First, subject-level GCNs (s-GCNs) model region-to-region structural connectivity within individual brains. Next, a long short-term memory (LSTM) network captures longitudinal trends in motor symptom progression. Finally, a population-level GCN (p-GCN) incorporates inter subject relationships to enhance predictions.
Evaluated on a cohort of 1,917 PD patients, the proposed model demonstrated superior performance compared to baseline models (e.g., SVM, KNN, Random Forest) and state-of-the-art (SOTA) graph-based models, including ST-GCN, BrainNetCNN, and EigenPoolingGCN. For current UPDRS-III predictions, the hierarchical GCN achieved a mean squared error (MSE = 3.24 ± 0.57) and R² = 0.91 ± 0.02, indicating high predictive accuracy with minimal error. For 3-year longitudinal predictions, the model achieved MSE = 4.78 ± 0.81 and R² = 0.87 ± 0.03, reflecting its ability to capture progressive changes in motor symptom severity. These results surpass the performance of competing models such as ST-GCN (MSE = 3.45 ± 0.53) and BrainNetCNN (MSE = 3.87 ± 0.62). These results demonstrate the superior capacity to model progressive changes in structural connectivity patterns and predict the trajectory of symptom severity with greater accuracy than existing models. The incorporation of GNNExplainer revealed biologically meaningful ROIs linked to motor symptom progression involving areas of the cortico-basal ganglia-thalamo-cortical loop, as well as the cerebellum, demonstrating the model’s interpretability. This hierarchical GCN framework represents a significant advancement in predictive modeling for PD, offering a robust, interpretable, and state-of-the-art tool for understanding motor symptom progression and guiding personalized clinical interventions. By combining static, temporal, and population-level insights, this approach provides a comprehensive analysis of PD progression, paving the way for future applications in neurodegenerative disease research.한국과학기술원 :바이오및뇌공학과
다중 단계 효소 반응의 수율 예측을 개선하기 위한 하이브리드 능동학습 전략
No full text학위논문(석사) - 한국과학기술원 : 바이오및뇌공학과, 2025.2,[iv, 37 p. :]We propose a hybrid active learning(HAL) framework to improve yield prediction in cell-free multi-step enzymatic reactions. Multi-step enzymatic reactions are critical in metabolic pathway optimization but are characterized by high experimental complexity and substantial time and cost requirements for data collection. To address these challenges, we utilize a query strategy that combines distance-based diversity with AI model prediction-based uncertainty to select the most informative experimental conditions, thereby maximizing the efficiency of AI model learning. The hybrid query algorithm demonstrated its effectiveness for efficient model learning with synthetic data from multi-step enzymatic reaction products, such as lycopene, butanol, and limonene, as well as in lycopene synthesis experiments. This approach highlights the potential for more efficient design of biological systems and optimization of metabolic pathways.한국과학기술원 :바이오및뇌공학과
하이프 사이클을 통한 네트워크 구조의 진화: 소셜 미디어에서의 AI Cloud 기대 분석
No full text학위논문(석사) - 한국과학기술원 : 기술경영학부, 2025.2,[iii, 55 p. :]This study investigates the alignment of public sentiment and network diffusion patterns of AI cloud technologies with Gartner's Hype Cycle by analyzing social media data. Using 20,272 tweets collected from 2019 to 2023, we conducted VADER sentiment and social network analyses to explore the evolution of public discourse. The results revealed that positive sentiment peaked during the "Innovation Trigger" phase and transitioned to more neutral and negative sentiments through the "Peak of Inflated Expectations" and "Trough of Disillusionment." Network analysis indicated a centralized structure dominated by key nodes in the early phases, evolving into a decentralized, community-oriented structure as the technology matured. The study validates the applicability of the Hype Cycle to social media data, offering a real-time framework for analyzing technology adoption. These findings contribute to innovation diffusion theory by linking network metrics, such as centrality and modularity, to stages of the Hype Cycle and provide actionable insights for strategists, policymakers, and marketers to tailor their approaches based on technology maturity.한국과학기술원 :기술경영학부
시장 위기 동안 스테이블코인과 기초 자산의 수익률 분석
No full text학위논문(석사) - 한국과학기술원 : 기술경영학부, 2025.2,[iii, 27 p. :]This study analyzes the returns of STEEM and Steem Blockchain Dollars (SBD), which are used within the Steem blockchain system, focusing on the ‘haircut’—–a mechanism intended to stabilize SBD’s value at the target price. Utilizing a theoretical model that links supply to return ratios, we analyze historical data of STEEM and SBD from September 2016 to March 2020. Our analysis identifies two significant haircut periods to evaluate the model’s accuracy against actual market behavior. Results indicate that during the first haircut period, SBD consistently outperformed STEEM, aligning with theory. In contrast, the second haircut period showed similar returns for both cryptocurrencies. This difference can be partly explained by the significant impact of the haircut period’s endpoint on return calculations, as slight shifts in timing can cause large variations in observed returns. Additionally, the perception of SBD as a stablecoin pegged to $1 may have limited its price volatility and reduced its theoretical advantage during the second haircut period. These findings reveal that stablecoin returns are influenced not only by supply mechanisms and market conditions but also by temporal factors and market participants’ perceptions. The study emphasizes the complexity of stablecoin dynamics in volatile environments and highlights the importance of careful timing and assumption selection in return analysis. Improving predictive models and incorporating broader market factors could offer a more comprehensive understanding of the mechanisms driving stablecoin behavior and better inform cryptocurrency investment strategies.한국과학기술원 :기술경영학부