https://scholar.gist.ac.kr/handle/local/7911 2026-09-30T17:23:01Z 2026-09-30T17:23:01Z MA, VAN LINH https://scholar.gist.ac.kr/handle/local/19889 2025-06-30T12:13:04Z 2024-12-31T15:00:00Z

Title: Visual Multi-Object Tracking with Re-Identification using Labeled Random Finite Sets Author(s): MA, VAN LINH Abstract: 본 논문은 베이즈 필터링(Bayesian filtering)을 통해 객체 출현 및 재출현 그리고 가림 현상과 같은 문제를 해결하는 향상된 온라인 비주얼 다중 객체 추적(multi-object tracking)을 기법을 소개한다. 먼저, 객체의 사라짐 및 재출현 문제 해결을 위해, 라벨링 된 확률 유한 집합(labeled random finite set) 필터링 알고리즘을 제안한다. 이는 객체 관리를위해객체들의특징을활용하고,객체검출(detection)에대해선형시간복잡도로 작동하는새로운모델을포함한다.또한,궤적의겹침과크기를고려한퍼지(fuzzy)검출 모델이 가림 현상(occlusion) 처리를 개선하며, 계산 시간을 최소화하기 위해 근사화된 필터도함께제시된다.다음장에서는제안하는다중객체추적알고리즘을 2D이미지에 서실제 3D좌표로확장한다.이접근법은단안카메라의 2D검출들을통합함으로써 3D 다중 객체 추적을 향상시키고, 카메라 재구성 시 검출기 재학습의 필요성을 제거한다. 이를통해트랙초기화및종료,재식별,가림현상처리를단일베이즈필터링재귀로통 합하고, 나아가 객체의 특징, 운동 데이터, 기하학적 가림 모델을 활용하여 제안 기법의 효율성을 높인다. 마지막으로, 2D 바운딩 박스 검출만을 사용하여 실시간 다중 카메라 다중 객체 추적(multi-camera multi-object tracking)을 위한 빠른 온라인 알고리즘이 고 안된다. 이 방법은 다중 센서(multi-sensor) generalized labeled multi-Bernoulli 필터를 – iii – 낮은복잡도와정확도손실없이구현가능하게하고,동적카메라구성에대한강건함을 보여준다.|This dissertation introduces advanced methods for online visual multi-object track- ing (MOT), addressing challenges like object appearance-reappearance and occlusion, using Bayesian filtering techniques. We first introduce an algorithm that leverages la- beled random finite set (LRFS) filtering to tackle disappearance and reappearance issues, incorporating a novel model that utilizes object features to manage reappearing objects with linear complexity relative to the number of detections. A fuzzy detection model is also introduced to enhance occlusion handling by considering track overlaps and sizes. To reduce computational time, we propose an approximation of this filter. In the next chapter, we extend our proposed multi-object tracking algorithm from 2D images to 3D real-world coordinates. More specifically, this approach enhances 3D multi-object tracking by integrating 2D detections from monocular cameras, elim- inating the need for detector retraining upon camera reconfiguration. This solution combines track initiation/termination, re-identification, and occlusion handling into a single Bayes filtering recursion, with improved efficiency through feature and kinematic incorporation, and a geometric occlusion model. In the last chapter, a rapid online al- gorithm is presented for real-time multi-camera multi-object tracking, using only 2D bounding box detections. This method simplifies the Multi-Sensor Generalized La- beled Multi-Bernoulli (MS-GLMB) filter to achieve a low-complexity implementation, demonstrating faster performance without accuracy loss, and robustness to dynamic camera configurations.

2024-12-31T15:00:00Z Hyun Myung Kim https://scholar.gist.ac.kr/handle/local/19883 2025-06-30T12:12:55Z 2022-12-31T15:00:00Z

Title: Vari-focal Light Field Camera for extended depth of field Author(s): Hyun Myung Kim Abstract: In recent years, as interest and demand for the metaverse and self-driving cars such as AR/VR have grown rapidly, great efforts have been made around the world to practicalize 3D depth sensing cameras. The light field camera, one of the three-dimensional depth-sensing cameras, is a camera that can detect depth with only one image sensor and is the most popular technology in practical aspects such as volume, cost, and battery consumption. Light field cameras can extract depth information from images such as those measured by multiple cameras, as microlens arrays act as key optical components for conventional cameras. In the existing light field camera, research on increasing the measurement range, angle of view, and depth resolution in the depth extraction area in the image processing area has been actively conducted. However, research in the hardware area of light field cameras was insufficient. Therefore, in this paper, we propose a study on the hardware implementation and characteristics of light field cameras. First, we propose an efficient processing method for microlens arrays, which are key components of light field cameras. Second, we devised a method of applying a variable-focus lens to widen the measurement distance of the light field camera. Third, we present a method for optimizing the light alignment of the microlens array to broaden the angle of view of the light field camera. Finally, we present a miniaturization method to apply light field cameras to mobile applications. Successful experimental demonstrations represent significant advances in the field of light field camera technology.

2022-12-31T15:00:00Z Daeyoung Han https://scholar.gist.ac.kr/handle/local/34611 2026-08-27T12:00:42Z 2025-12-31T15:00:00Z

Title: Uncertainty-Driven Solution Space Exploration for Single-Image Super-Resolution Author(s): Daeyoung Han Abstract: Single image super-resolution (SISR) is inherently ill-posed: a single low-resolution observation is consistent with many distinct high-resolution reconstructions. Geometrically, the plausible reconstructions form a solution space, the intersection of the feasible set defined by the degradation operator and the manifold of natural images, whose width varies dramatically across the image. Smooth regions correspond to a near-point in this space, while edges and textures correspond to a much wider set of admissible alternatives. Existing generative SISR methods explore this space without guidance: they draw stochasticity from isotropic priors that treat all pixels as equally ambiguous. This dissertation argues that such guidance can be estimated directly from the low-resolution input. A theoretical analysis identifies the aleatoric uncertainty of the SISR problem with the magnitude of the residual between the ground truth and the prediction of a pixel-wise-trained baseline. Geometrically, this identity says that the uncertainty map is the local width of the solution space. A frozen baseline together with a learned uncertainty estimator therefore yields a residual parametrization in which the centroid of the space and stochastic displacements around it are explicitly separated. Two frameworks exploit this parametrization. The first explores the solution space at inference time, using the uncertainty map as the covariance of an anisotropic prior over residuals so that different forward passes produce different points in the space. The second explores it at training time: a learned synthesizer generates probe samples at calibrated distances from the natural image manifold, and these probes drive a contrastive minimax game in which the discriminator learns the geometry of the solution space directly. Together, the two frameworks demonstrate that aleatoric uncertainty is the natural coordinate system for the solution space, and that exploring this space along its native coordinates yields a more favorable perception-distortion trade-off across standard benchmarks.|단일 영상 초해상도(Single Image Super-Resolution, SISR)는 본질적으로 잘못 설정된 문제(ill-posed problem)이다. 하나의 저해상도 입력은 다수의 서로 다른 고해상도 복원 결과와 부합할 수 있기 때문이다. 이러한 가능한 복원 결과들은 기하학적으로 해 공간(solution space)을 이룬다. 해 공간이란, 열화 연산자에 의해 정의되는 가능 집합과 자연 영상 다양체의 교집합으로서, 그 폭은 영상 내 위치에 따라 크게 달라진다. 평탄한 영역은 해 공간이 거의 한 점으로 수렴하는 반면, 윤곽선이나 질감 영역은 훨씬 더 넓은 후보 집합을 갖는다. 그러나 기존의 생성 모델 기반 초해상도 기법들은 이러한 해 공간의 구조를 인지하지 못한 채 탐색을 수행한다. 모든 픽셀을 동일한 정도로 모호하다고 가정한 등방성(isotropic) 사전 분포로부터 무작위성을 도입하여 학습할 뿐이다. 본 학위논문은 이러한 해 공간을 탐색하기 위한 단서를 저해상도 입력으로부터 직접 추정할 수 있음을 주장한다. 이론적 분석을 통해, 초해상도 문제의 우연적(aleatoric) 불확실성이 픽셀 단위 손실로 학습된 기준 모델의 출력과 정답 영상 간 잔차의 크기와 정확히 일치함을 확인하였다. 이를 기하학적으로 해석하면, 추정된 불확실성 맵은 해 공간의 국소적 폭을 의미한다. 따라서 사전 학습된 기준 모델과 학습 가능한 불확실성 추정 모델을 결합하면, 해 공간의 중심과 그 주변에서의 확률적 변위를 명시적으로 분리하는 잔차 매개변수화(residual parametrization) 방법을 활용할 수 있다. 본 연구에서는 이러한 매개변수화를 활용하는 두 가지 프레임워크를 제안한다. 첫 번째 프레임워크는 추론 단계에서 해 공간을 탐색한다. 불확실성 맵을 잔차에 대한 비등방성(anisotropic) 사전 분포의 공분산으로 사용함으로써, 매 추론마다 해 공간 내 서로 다른 지점이 생성되도록 한다. 두 번째 프레임워크는 탐색을 학습 단계에서 진행한다. 학습된 합성 모델이 자연 영상 다양체로부터 일정 거리에 위치한 탐색용 표본(probe sample)을 생성하고, 이 표본들을 활용한 대조적(contrastive) 미니맥스 게임을 통해 판별자가 해 공간의 기하 구조를 직접 학습하도록 유도한다. 두 프레임워크는 우연적 불확실성이 초해상도 해 공간의 자연스러운 좌표계임을 보이며, 이를 따라 해 공간을 탐색하는 것이 다양한 표준 벤치마크에서 더 유리한 화질-왜곡 절충을 달성함을 입증한다.

2025-12-31T15:00:00Z Pavel S. Ni https://scholar.gist.ac.kr/handle/local/33277 2025-12-30T12:07:42Z 2020-12-31T15:00:00Z

Title: Ultrasound Medical Imaging: A Novel Super-Resolution Imaging Technique Utilizing Random Interference Author(s): Pavel S. Ni Abstract: Medical imaging modalities are used every day in hospitals to examine medical condi-tions, consequently allowing us to analyze anatomical, physiological, metabolic, and functional information of the human body. Ultrasound is one of the most widely used diagnostic tools be-cause it is affordable and non-invasive. Ultrasound is used to assess the tissue, vessels, and or-gans within the human body. In the past few decades, a lot of effort was put by the academic community to improve the quality of ultrasound images. However, despite decades of innova-tion, the main disadvantage of ultrasound is low image resolution. In this dissertation, we develop a novel super-resolution imaging technique utilizing constructive and destructive interference of ultrasonic waves. In particular, we first introduce a method to generate an incident ultrasonic wavefront of random interference. This wavefront then has a spatially variant property that yields individual spatial points, in the region of interest, to reflect mutually incoherent spatial impulse responses. Second, we develop an image recon-struction method based on an L1-norm minimization algorithm that is capable of identifying the scattering points by the presence of spatial impulse responses in the received echo signals. The natural synergy between the properties of the wavefront of random interference and the image reconstruction algorithm allowed us to create the necessary conditions for a successful recon-struction of super-resolution ultrasound images. Lastly, we demonstrate using numerical simula-tions and phantom experiments that the proposed method can achieve four times better spatial resolution. In the simulation study, the proposed method achieved a resolution of 0.25 mm. In the real phantom experiment, we demonstrated that the proposed method can successfully re-construct ultrasound images of nylon wires as small as 0.08 mm in diameter using a tissue-mimicking phantom. We argue that the proposed method is a big step towards achieving super-resolution ultrasound imaging and offers a new perspective on ultrasound imaging. The pro-posed imaging method bypasses the diffraction resolution limit by eliminating the need for the conventional focused ultrasound beam.

2020-12-31T15:00:00Z