지역 탐색을 위한 복수 무인기의 임무 분할 및 제어를 위한 지상관제시스템 설계 및 구현

이병욱 韓國外國語大學校 大學院 2018 국내석사

Multiple UAVs are used for surveying and measuring on an area observation missions. The existing mission executing to using multiple UAV is focused on formation flight to enhance the stability of UAV mission executing. When the stability of mission performance of a single UAV is increased due to the development of UAV, it is more efficient for the multiple UAVs to perform the division mission rather than performing the observation mission and survey mission by formation flight. In this paper, we propose a design of the ground control system that enables efficient communication without overhead between multiple UAVs and the ground control systems, and propose a script for mission division. Also it describes the structure of the ground control system and the method of division of missions that can efficiently perform area survey using multiple UAVs when stability of mission performance of UAV is enhanced in the near future.

자율주행 로봇을 위한 다중 특징 기반의 외부장면 이해

김대년 울산대학교 2010 국내박사

This dissertation describes a method for the scene understanding of the intelligent robot systems (IRS) on outdoor environment. To understand the scene, the IRS can perform several functions such as image segmentation, object detection, landmark detection, recognition, localization, navigation, object analysis and so on. To identify objects, the robot recognizes the object from an image taken by moving robot on outdoor scene. As a beginning, this dissertation presents the candidates for a segment of region to building of artificial object, sky and trees of natural objects. Then we define their characteristics individually. In the process, we segment the regions objects included by preprocessing using multiple cues. Multiple cues are HSI (Hue, saturation, intensity), line segments, context information, HCM (Hue Co-occurrence Matrix), PCs (Principal Components) and vanishing point. So, we show the combination of appearance and context information. Detection of buildings is considered as the best landmark with dense appearance in the city. The outer surface of building comprises of some special properties of manmade objects such as rectangular shape, wall region, windows, entrance and columns. These characters support the information for classifying the buildings with other objects and identifying to each other. The buildings are classified with other objects like sky, trees, bushes and roads. The building comprises of three major parts as a hierarchical system for understanding of intelligent robot systems. Firstly, line segments and two neighborhood regions are extracted. A model of line segment (MLS) is constructed by color information of neighborhood regions. MLS is used to reduce the line segments of non-building patterns. Secondly, the rest of line segments are clustered such parallel lines which have a common vanishing point (VP) by MSAC (m-estimator sample consensus) algorithm. The maximum numbers of VPs calculated for vertical and horizontal directions are one and five, respectively. The vertical and one of horizontal clusters create a mesh of convex quadrangles (skew parallelograms) as a candidate face of building. The geometrical properties like distributed density of line segments and number of intersections are analyzed and considered as criteria to refine the building face. Finally, the building facets detect and represents by a boundary quadrangle with its area and vertical and horizontal VPs. The second part is for identifying the building to each other by a test image. The test image is taken when the robot is working. Then it is matched against the stored images in a database. To do so, wall region of building facet is extracted. Then a wall color histogram is calculated for each facet. The matching process contains this step. In this step, by using the wall color histogram and the area of a test facet, the small number of the stored facets in the database is chosen as a set of candidates (sub-candidates). Here, we proposed a new method, cross ratio-based refinement, for exactly verifying the correspondences of image pairs. With the learned database and exact recognition process, the IRS can localize and navigate in the environment. In the last part, the geometrical information of building's principal components (PCs) such as wall regions, windows and entrance is analyzed in more detail. Here, the PCs are detected and then the edges and lengths of windows are accurately estimated. The property estimates the number of story and rooms in the same floors by merging skewed parallelograms of the same color. The relative geometrical information of PCs like quantity, height and alignment are computed. The geometrical information is available for scene with 3D model of building so that the system understands a more in semantics of the environment where IRS explores. Finally, the analytical experiments analysis and present the results of outdoor scene understanding using multiple cues and the geometrical properties analysis. Extensive experiments with the object segmentation, recognition and analysis on outdoor scene confirm the validity of the approach. 본 논문은 지능로봇시스템 (intelligent robot system)이 외부환경의 장면 (scene)을 이해하는 방법에 대해서 기술한다. 외부환경의 장면을 이해하기 위해서, 지능로봇시스템 (IRS)은 물체검출 (object detection), 영상분할 (object segmentation), 표식검출 (landmark detection), 인식 (recognition), 위치추정 (localization), 주행 (navigation), 물체분석 (object analysis)과 같은 여러 기능 등을 실행할 수 있다. 물체를 확인하기 위해서, 로봇은 외부환경의 장면에서 주행하는 로봇에 의해 물체를 인식한다. 먼저, 본 논문은 자연물의 하늘 및 나무, 인공물의 빌딩에 대해서 영역을 분할하는 것을 나타낸다. 영역을 분할할 때 우리는 물체의 특성을 개별적으로 정의한다. 이 과정에서, 다중 큐를 이용하여 전처리에 의해 포함된 영역의 물체를 분할한다. 다중 큐는 색상 (color), 선분 (line segment), 상황 정보 (context information), 색상 동시발생 행렬 (HCM), 주요한 요소성분 (principal components) 및 소실점 (vanishing points)을 이용한다. 이렇게, 물체의 분할은 외관 (appearance)과 상황정보의 조합을 보여준다. 빌딩의 검출은 도시에 있는 가장 조밀한 외관을 가진 표식으로 고려된다. 빌딩의 벽면은 직사각형 모양의 벽 영역, 창문 및 현관과 같은 사람에 의해 만들어진 특별한 속성을 가진다. 이런 속성은 다른 물체와 빌딩을 분류하고 확인을 하기 위한 정보로서 제공한다. 빌딩은 하늘, 나무, 숲 및 도로 같이 다른 물체로 분류된다. 빌딩은 지능로봇시스템의 이해를 위한 계층적인 시스템으로 세 부분으로 구성한다. 첫 번째로, 선분과 두 이웃 영역을 추출한다. 선분의 모델 (model of line segment)은 이웃 영역의 색상 정보에 의해 구성한다. 선분의 모델 (MLS)은 빌딩이 아닌 곳의 선분을 제거하기 위해서 이용된다. 둘째로, 선분의 나머지는 MSAC (m-estimator sample consensus) 알고리즘에 의하여 일반적인 소실점 (VPs)이 있는 평행선에 밀집한다. 수직과 수평선분의 방향에 대해서 소실점의 최대 다섯 개를 산출한다. 수평과 수직선분의 교차로 빌딩 벽면의 후보인 평행사변형의 그물을 만든다. 분포된 선분의 조밀도와 같은 기하학적인 속성과 교차점의 수는 빌딩면을 세밀하게 구분하는 기준으로서 고려하고 분석한다. 마지막으로, 빌딩면은 사각형의 영역, 경계선과 수직, 수평선분의 소실점에 의해 검출된다. 두 번째는, 테스트 영상에 의하여 각각 다른 빌딩을 확인한다. 테스트 영상은 로봇이 주행할 때 이용한다. 그때 데이타베이스에 저장되어 있는 영상과 일치함을 확인한다. 이렇게 하여, 빌딩면의 벽 영역은 추출된다. 이때 벽 색상 히스토그램 (wall color histogram)은 각 빌딩면에 대해 산출한다. 매칭 과정은 이 단계를 포함한다. 이 단계에서, 벽 색상 히스토그램과 테스트 빌딩면의 영상 영역을 이용하여 데이타베이스에 저장된 작은 수의 빌딩면은 후보의 집합으로 선택된다. 여기에서, 우리는 정확하게 두 영상의 확인을 위한 종횡비를 기반하여 미세하게 구별을 하는 새로운 방법을 제시했다. 데이터베이스에 대한 학습과 정확한 인식 과정으로, 지능로봇시스템은 외부환경에서 위치를 인식하고 주행할 수 있다. 마지막은, 벽 영역, 창문과 정문과 같은 빌딩의 주요한 요소성분의 기하학 정보는 더 자세히 분석된다. 여기에서, 빌딩의 주요한 요소성분은 검출되고 창문의 에지와 길이는 정확하게 추정된다. 빌딩의 주요한 요소성분의 속성은 동일한 색상의 평행사변형을 합병해서 빌딩에서 층 (story)의 수와 동일한 층 (floor)에 있는 방 (rooms)의 수를 추정한다. 빌딩의 주요한 요소성분과 상대적인 기하학 정보는 층의 높이와 방의 정렬과 같은 양으로 계산된다. 기하학 정보는 지능로봇시스템이 주행하는 환경의 의미론에서 더 많은 것을 이해하기 위해 빌딩의 삼차원 모델에 대한 장면을 이용한다. 마지막으로, 분석에 입각한 실험분석은 다중 큐 및 기하학적인 속성을 분석하여 외부장면의 결과를 나타낸다. 외부환경의 장면에서 물체 분할, 물체인식과 물체분석에 대한 상세한 실험은 접근의 타당성을 확인한다.

드론 및 보트를 활용한 수질 감시 미션 협업 시스템

조용아 韓國外國語大學校 大學院 2019 국내석사

Robotic systems are used in various environments (nuclear power plants, deep sea exploration, Mars exploration, etc.) that are given a mission from the ground control system and are difficult for human access. Recent missions are becoming more difficult to perform with the resources given to a robot (such as limited battery and available equipment limitations) as their types and environments evolve complexly. For this purpose, there are many previous researches on multiple homogeneous robots that divide a given mission, but it is difficult to apply it to a mission which is difficult to solve by homogeneous robots alone. In this paper, we design a mission collaboration system for a water quality monitoring mission by designating a heterogeneous robot as a drone and a boat to design a collaborative mission execution system between heterogeneous robots.

치매 진단을 위한 Faster R-CNN 활용 MRI 바이오마커 자동 검출 연동 분류 기술 개발

손주형 韓國外國語大學校 大學院 2019 국내석사

In order to diagnose and prevent Alzheimer's Disease (AD), it is becoming increasingly important to develop a CAD(Computer -aided Diagnosis) system for AD diagnosis, which provides opti mized treatment for patients by analyzing 3D MRI images. It is essential to apply powerful machine learning such as deep learn ing in order to automatically classify stages of Alzheimer's Dise ase and to develop a Alzheimer's Disease support diagnosis sys tem that has the function of detecting hippocampus and CSF w hich are important biomarkers in diagnosis of Alzheimer's Disea se. In this paper, we classify MRI cases into three categories of AD, mild cognitive impairment, and normal control according t o MRI image by applying 3D brain MRI image to the Faster R-CNN model and detect hippocampus and CSF in MRI image in real time do. To do this, we use the 2D MRI images extracted from the 3D MRI images of the Faster R-CNN, and perform th e widly used Majority Vote on the resulting bounding box label s. To verify the proposed method, we used the ADNI data set, which is the standard brain MRI database. Experimental results show that the proposed method achieves improved classification performance compared with other state-of-the-art methods.

얼굴인식 성능 개선을 위한 얼굴 전역 및 지역 특징 기반 앙상블 압축 심층합성곱신경망 모델 제안

윤경신 韓國外國語大學校 大學院 2020 국내석사

In this paper, we propose a novel knowledge distillation algorithm based on combination of local features and global features of face images. The aim of our work is to build a compressed ensemble DCNNs that are very similar in face recognition performance to deep ensemble networks. The purpose of knowledge distillation is to transfer the information of recognition result of the ensemble network to a single neural network, dramatically reducing recognition execution time and ensemble neural network storage space while maintaining high recognition performance. In order to transfer the capability of high-level recognition performances of the ensemble neural networks to a single neural network, the probability for class prediction, which is the softmax output of the ensemble neural network, is used as a soft target for training a single neural network. By applying the knowledge distillation algorithm, the local feature informations obtained by training the ensemble deep neural network using facial subregions of the face image as input are transmitted to a single deep neural network to create a so-called compressed ensemble deep neural network proposed by us. The compression model can considerably improve face recognition performance by simultaneously learning the local and global features of facial expressions. The experimental results demonstrate that our proposed compressed ensemble deep network can maintain the recognition performance of the ensemble deep network and is superior to the recognition performance of a single deep network. In addition, our proposed method can significantly reduce the storage (memory) space and execution time, which is main bottleneck for the use of ensemble deep networks due to a large number of parameters (usually, millions). Considering both effectiveness and efficiency, our proposed compressed ensemble deep network might be applied for practical face recognition applications with high-level recognition accuracy.

트랜스포머를 이용한 전동기 전류 데이터 이상 탐지 및 이상 상태 분류

송영종 韓國外國語大學校 大學院 2023 국내석사

전동기는 산업, 가전, 운송 등 다양한 분야에서 사용된다. 전동기의 고장은 막대한 경제적 손실을 초래하고, 특히 운송에 사용 되는 전동기는 막대한 인적 피해를 초래할 수 있기 때문에 전동기의 이상을 감지하는 것은 매우 중요하다. 딥러닝을 이용한 이상 탐지 연구가 활발히 진행되고 있으며, 실제 현장에서 이상 데이터를 획득하기 어려워 정상 데이터만 훈련하고 이상을 탐지하는 비지도 학습 기반 이상 탐지 방법이 주로 연구되고 있다. 본 논문에서 사용된 데이터 세트는 12개의 전력을 가진 40개의 전기 모터 인버터의 RST 3상 전류 센서 데이터로 구성된다. 정상 데이터와 전동기에서 주로 발생하는 4가지 이상 상태(베어링 불량, 회전체 불균형, 벨트 느슨함, 축 정렬 불량) 데이터로 구성된다. 본 논문에서는 트랜스포머 모델을 사용하여 비지도 학습 기반 이상 탐지 및 지도 학습 기반 이상 상태 분류 방법을 제안한다. 첫째, 비지도 학습 기반 모델을 사용하여 이상 탐지한 다음, 지도 학습 기반 모델을 사용하여 이상 상태를 분류한다. 이상 상태 분류 성능은 한 모델이 데이터 세트의 모든 모터인 12개 전력을 가진 40개 전기 모터 센서 데이터의 4개 이상 상태를 분류할 때 0.94AUC 점수로 매우 강력하다.

Breast cancer classification into molecular subtypes using 3D MRI

Naima Nasir Hankuk University of Foreign Studies, Graduate Sch 2018 국내석사

The classification of breast cancer into molecular subtypes holds a lot of significance for appropriate and effective treatment at various stages. MRI (Magnetic Resonance Imaging) is an efficient and relatively safer method to capture different organs in the body, aiding in diagnosis of the disease and staging information. Additionally, CAD (Computer Aided Diagnosis) make the process easier by marking the malignant region or tumor in the organ. 3D MRI breast scans integrated with CAD information are an effective way to diagnose the tumor and the staging of cancer. Previous classification approaches mainly focused on clinical data/genomic info (expensive biopsy test etc.). In this paper, we propose to classify cancer based on pure image-based knowledge embedded in the MRI image. We propose to use two effective texture features, i.e., the LBP local binary patterns and the Gabor Square Energy, to learn a conventional discriminative model. We also present an appropriate deep CNN (convolutional neural network) architecture to enable classification directly from a raw tumor volume, (reducing the high variance/overfitting problem). Both the approaches are evaluated on classifying luminal A and others, whereas, the CNN-based approach is evaluated for classifying luminal A, B, and HER2. Comparing to the previous image-based classification approaches, the proposed approach showed a noteworthy improvement in classification. 분류하는 것은 다양한 단계에서 적절하고 효과적인 치료법에 많은 중요성을 가진다. MRI (Magnetic Resonance Imaging)는 신체의 여러 장기를 포착하여 질병 및 스테이징 정보의 진단을 돕는 효율적이고 비교적 안전한 방법이다. 또한 CAD (Computer Aided Diagnosis)는 장기의 악성 영역이나 종양을 표시함으로써 과정을보다 쉽게 만든다. CAD 정보와 통합 된 3D MRI 유방 스캔은 종양 및 암 병기를 진단하는 효과적인 방법입니다. 이전의 분류 방법은 주로 임상데이터 / 게놈 정보 (고가의 생검 검사 등)에 중점을 두었습니다. 본 논문에서는 MRI 영상에 내장 된 순수한 영상 기반의 지식을 기반으로 암을 분류하는 방법을 제안한다. 종래의 식별 모델을 학습하기 위해 로컬 바이너리 패턴과 (LBP) Gabor 스퀘어에너지와 같은 2 개의 효과적인 텍스처 피쳐를 사용하는 것을 제안한다. 또한 원시 종양 볼륨에서 직접 분류 할 수있는 적절한 CNN (Convolutional Neural Network) 아키텍처를 제시한다 (High variance/overfitting 문제 감소). 두 접근법 모두 luminal A와 others를분류 할 때 평가되는 반면, CNN 기반 접근법은 luminal A, B 및 HER2를 분류하기 위해 평가된다. 이전의 이미지 기반 분류접근법과 비교하여, 제안 된 접근법은 분류에서 현저한 개선을 보였다.

미세먼지 농도 예측을 위한 어텐션 기반 합성곱 장단기 메모리 모델

이준민 한국외국어대학교 대학원 2023 국내석사

Air pollution, particularly fine dust, poses a significant threat to public health and necessitates accurate prediction models for effective mitigation strategies. In this paper, we propose an attention-based ConvLSTM-DNN networks for fine dust prediction. Our model integrates the feature extraction capabilities of a 2D Convolutional Neural Network (CNN) with the long-term memory retention of an LSTM, capturing spatial and temporal dependencies in the input data. We introduce an attention mechanism to enhance the model's predictive ability by assigning greater weight to influential regions or features. This mechanism improves overall accuracy. Additionally, we optimize pre-processing of time-series data to maximize performance. Using real-world weather and fine dust concentration data, we conduct experiments to evaluate our model. Comparative analyses with existing models demonstrate its superior prediction accuracy. Our study has implications for air quality management and public health. The proposed model offers accurate and reliable fine dust forecasts, aiding decision-making processes. By leveraging ConvLSTM, attention mechanisms, and optimized pre-processing, our model advances air pollution monitoring and control. In summary, our paper presents an attention-based ConvLSTM-DNN networks for fine dust prediction. Experimental results highlight its superiority and potential in improving air quality and safeguarding public health.

비디오 안면 인식 성능개선을 위한 하이브리드 심층합성곱신경망 결합 구조 개발

김경태 韓國外國語大學校 大學院 2019 국내석사

In this paper, we propose a novel way of combining multiple deep convolutional neural network (DCNN) architectures which work well for accurate video face recognition (FR) by adopting a serial combination of 3D and 2D DCNNs. The proposed method first divides an input video sequence (to be recognized) into a number of sub-video sequences. The resulting sub-video sequences are used as input to the 3D DCNN so as to obtain the class-confidence scores for a given input video sequence by considering both temporal and spatial face feature characteristics of input video sequence. The class-confidence scores obtained from corresponding sub-video sequences is combined by forming our proposed class-confidence matrix, which is then used as an input for learning 2D DCNN learning which is serially linked to 3D DCNN. Finally, fine-tuned, serially combined DCNN framework is applied for recognizing the identity present in a given test video sequence. To verify the effectiveness of our proposed method, extensive and comparative experiments have been conducted to evaluate our method on COX face databases and YTC databases with their standard face identification protocols. Experimental results showed that our method can achieve better or comparable identification rate compared to other state-of-the-art video FR methods.

모션 스테레오에 의한 단일 카메라 기반의 계단 검출 : STAIRWAY DETECTION BASED ON SINGLE CAMERA BY MOTION STEREO

카세레스 헤르난데스 다닐로 울산대학교 2011 국내석사

Ad-Hoc localization is a fundamental task in computer science, which means imposing some physical order upon a set of objects, so that a given object has a greater probability of being in some particular regions of space than in others. In other words, localization is the process by which the system determines its pose with respect to object on the basis of sensory measurements. In this dissertation, we ropose to detect the localization and recognition of indoor stairways. This is a fundamental step for the implementing of autonomous stair climbing navigation, as well as the implementing of passive alarm systems intended for blind and visually impaired people. Both of these systems must be able to recognize parameters that can describe stairways in unknown environments. The proposed method analyses edges of the stairway based on planar motion tracking and directional filters. First, we extract frames in an interval that started just before the earliest detection from the video capture sequence. Second, based on the assumptions that indoor stairways are located below of the true horizon, we propose to nd the distance from the camera to the true horizon or maximum distance of the ground plane (MDGP) from the sequence of images. Third, we extract the edge line segments of the stairs by using the Gabor filter, and the resulting set of line segments will be considered as a candidate of stair-step. Fourth, using the set of line segments as our region of interest (ROI) we extract the hypothetical set of targets from the sequence image at time t and t +1, respectively. Fifth, we find the optimal target correspondence between the frames at times t and t + 1 by using the correlation method. Sixth, in determining the distance discrepancy we propose a discrimination method for ground plane. Consequently, the remaining information is considered as an indoor stairway candidate region. As a result, testing was able to prove its effectiveness.