Visual Servo Control of a Bolting Robot for Automatic Construction of Steel Structure using Camshift

Choe, Yun Geun 고려대학교 대학원 2010 국내석사

In order to use a robot in the construction automation filed, we proposed the concept of Bolting Robot and a visual servo control scheme to track a bolting tool to a bolt-hole in the structural steel frame. To track a bolting tool to a bolt-hole, new visual servo control scheme was proposed in this thesis. It compromises tracking part, image preprocessing, and CHT to estimate a bolt-hole location. Tracking part enables to track the selected bolt-hole quickly. And tracking part can give some information about bolt-hole location such as a radius of the bolt-hole, a center position of the bolt-hole, and a position of region of interest (ROI). This information helps to reduce image processing time because image processing region can be reduced. For estimating a location of a bolt-hole, Circular Hough Transform (CHT) was used to extract circles. Generally, CHT is computationally complex because the search program is to find parameter triplets (a, b, r) of circles. A distance from a camera to a steel frame can be measured by using laser range-finder installed. The radius of a bolt-hole can be calculated with the distance to a steel frame. Since the radius is known, the processing of CHT can be reduced to 2D. In addition, it contains image pre-processing to make an image of bolt-holes to be clear. Pre-processing has 3 steps which consist of compensating lens distortion, noise filtering, and edge detection. By the proposed visual servo control scheme, we have tested that our bolting robot inserts a bolt to a bolt-hole. It can be successfully run with this resolution.

Vehicle side-view detection by ordered independent parts with convolutional neural network

Park, Richard Young Graduate School, Korea University 2016 국내석사

In this thesis, a scheme to improve detection capability of vehicles by their sideviews with the vehicle detection by ordered independent parts (VDOIP) algorithm, which is based on existing vehicle detection by independent parts (VDIP), is proposed. By detecting the moving directions of the font and rear parts of vehicles, and using logical pattern matching between the parts based on their orders or relative locations in the scene, the proposed scheme can detect the vehicles as single pieces more e↵ectively. The capability and the performance of the proposed VDOIP method were confirmed using video clips of real street scenes.

Pattern matching for object recognition and registration using geometry based vector mapping descriptor

유용탁 Graduate School, Korea University 2017 국내석사

Object recognition has always been a troublesome issue for computer vision. Despite continuous researches, it still remains a challenge to define features, match the corresponding features and develop accuracy and precision concurrently with considering speed and robustness at the same time. In this thesis, we propose a novel feature matching method called Vector Mapping Descriptor (VMD) to overcome existing issues. We implement the sub-pixel units for edge detection to improve the accuracy of invariant features. Then sub-pixel unit edges are enhanced by least square error estimation. More accurate geometric features are extracted from enhanced sub-pixel unit edges. The geometric features are extracted from the geometric shape of the object. We defined two geometric features which are center of circle and intersection of lines. The geometric features are used to construct VMD. The VMD represents the correlation of features that consist of the Euclidean distance and angle. The geometry based VMD for pattern matching is proposed to effectively match salient feature points between different images under geometric transformation irrelevant to missing or additional feature points. The VMD induces the one to one corresponding feature matching. The group of one to one corresponding feature points from different images results in the completed object matching. The proposed matching algorithm is invariant to geometric transformation such as translation, rotation, and scaling differences. It can also cope with the partial distortion and occlusion. The experiments are conducted with industrial camera to show that our system can be executed in real-time.

Improving position estimation algorithm using vehicle radar

원종영 고려대학교 공학대학원 2017 국내석사

ADAS(Advanced Driver Assistance System)는 차량 주행 시 사고를 미연에 방지하기 위한 능동 안전 시스템을 말하며 다양한 기능으로 구성되어 있다. 그중 레이더 시스템은 원거리의 장애물을 감지하여 실시간으로 차량을 제어하기 위해서 필요한 핵심 기능이다. 차량 레이더 시스템에서 표적을 탐지하고, 추적하기 위해서 위치 추정은 필수적인 기능이며 칼만 필터가 보편적으로 사용되고 있다. 하지만 칼만 필터는 시스템 모델링의 오류, 측정 잡음과 시스템 잡음의 변동에 민감하고 측정치가 급격하게 변화할 경우 추정값의 정밀도가 떨어진다. 그에 반해 FIR 필터 (Finite Impulse Response Filter)는 측정 잡음과 시스템 잡음 변동에 강인하며 급격한 측정값 변동에도 안정적으로 추정치를 얻을 수 있으므로 칼만 필터의 단점을 보완할 수 있다. 본 논문에서는 차량 레이더를 활용해 취득한 데이터를 활용하여 주행 상황별로 칼만 필터와 FIR 필터 파라미터값을 변경하면서 각 위치 추정치의 정확도를 실험한다. 이를 통해 칼만 필터 대비, FIR 필터의 파라미터 오차 강인성을 확인하고, FIR 필터의 위치 추정치와 칼만 필터의 위치 추정치를 비교 분석하여 특정 주행 상황에서 FIR 필터의 유용성을 입증한다.

Advanced template matching algorithm using geometric component

김재현 고려대학교 공학대학원 2017 국내석사

컨베이어 벨트 위의 물체를 촬영한 이미지에서 물체의 위치 파악은 매우 중요하다. 랜덤하게 올려져 있는 컨베이어 벨트의 물체를 피킹(Picking)하기 위함이다. 이미지에서 대상 물체의 위치와 회전 각도, 물체의 수를 찾는 것은 영상 인식의 중요한 분야이고, 많은 연구가 진행되고 있다. 하지만 기존에 널리 사용되고 있는 NCC(Normalized Cross Correlation) 기반의 매칭 방식은 크기 변화에 대해 오차를 많이 발생시킨다. 이 때문에 컨베이어 벨트에 크기가 다른 물체가 섞이거나, 다른 물체에 올라탄 물체가 원근법에 의해 크기가 달라져 보인다면 NCC 방식의 매칭 결과는 오차가 발생한다. NCC 방식의 알고리즘은 모델 이미지 그대로 콘볼루션 연산을 수행하기 때문에 고정된 크기로만 매칭이 가능하여 크기 변화에 약점을 갖는다. 모서리 특징점 추출 방식은 명확한 매칭을 위해 사용되는데, 대상 물체의 크기 변화에는 이 특징점들이 균일하게 상대적으로 변화된다. 따라서 본 논문에서는 상대적인 검출 방식인 구성요소 방식과 기하학적 특징점을 이용하는 기하 구성요소 기반 매칭방식(Geometric Component Based Matching)을 제안한다. 제안하는 알고리즘의 성능을 검증하기 위해 MIL(Matrox Image Library)의 기하 모델 검출(Geometric Model Finder)방식과 OpenCV의 NCC패턴 매칭에 대하여 물체의 좌표, 성공률, 정확도의 3가지 항목을 비교하였다.

Modified discrete gaussian-hermite moment에 기반한 향상된 SURF 알고리즘

백관룡 高麗大學校 大學院 2012 국내석사

In this dissertation, we introduce a modified Discrete Gaussian-Hermite Moment (MDGHM) and propose MDGHM-based SURF algorithm (MDGHM-SURF). The MDGHM is a movable mask form of DGHM. It has controllable sampling intervals that allows the local features of a non-square image to be represented. Therefore, MDGHM can be applied to non-square images without any image truncation. The MDGHM is applied to the SURF algorithm in the describing interest points instead of Haar wavelet responses. Evaluation with previous algorithms demonstrate that the proposed method is more robust to variations such as scale, rotation, illumination, blur, JPEG-compression and viewpoint in feature extraction.

템플릿 매칭과 타원 근사법을 이용한 강인한 볼트구멍 인식 시스템

임종욱 高麗大學校 大學院 2011 국내석사

Recently, the automation for constructing a building has been extensively studied. The bolting robot among the various robots for construction such as a fireproofing spray robot and a tool developing robot runs to assemble bolts to holes of the steel frame in the construction fields. When the bolting robot assembles the bolt, the most important work is to distinguish the bolt hole from the steel frame. This thesis suggests the method to detect the bolt hole of the steel frame using the template matching, the circular Hough transform and the ellipse fitting. The template matching can be used to roughly estimates the location of the object in the various illumination. The circular Hough transform can be used to discover more accurate location of the object than the template matching. The ellipse fitting can reduce the error of the circular Hough transform. Therefore, we compare the performance between the each method and the mixed methods. We confirm that the proposed method is environment robust.

Multi-target tracking algorithm based on FIR filters

민경민 Graduate School, Korea University 2014 국내석사

In this thesis, we introduce a minimum variance finite impulse response (MVFIR) filter and propose a new multi-target tracking algorithm based on FIR filter in framework for both linear system and linear jump Markov system. The FIR filter estimates the state of system from recent input and output information. The advantage of FIR filter is robustness to effect for both model uncertainty and incorrect noise. Contrastively, a filter based on infinite impulse response (IIR) structure may perform poorly or divergence in such effects. Therefore, multi-target tracking problem with model uncertainty or incorrect noise can be solved by FIR filtering algorithm. The proposed algorithm conducts estimations of all possible states of the targets by MVFIR filtering from the measurements, which received from a sensor involving clutters. All possible combinations to make vectors of measurements on the recent time interval are constructed and conducted by FIR filtering. To extract the real state of the targets, likelihood function in a batch form of system model and the Mahalanobis distance are used. To evaluate the proposed algorithm, the simulation is conducted with Gaussian mixture probability hypothesis density (GM-PHD) filter. The optimal subpattern assignment metric and the error of estimated number of target are used as performance indicators.

Obstacle avoidance path planning based on output constraints model predictive control

김지창 Graduate School, Korea University 2019 국내석사

Image processing and control technologies have been widely studied and autonomous vehicles have become an active research area. For autonomous driving, it is essential to generate a safe obstacle avoidance path considering the surrounding environment. This paper devised an algorithm based on a real-time output constraints model predictive control for obstacle avoidance path planning in high speed driving situations. The proposed approach was compared with the normal predictive control approach by simulation, including operation times to verify robustness for high speed driving situations. We used the ISO 2631-1 comfort level standard to quantify driver comfort for both cases.

강화학습을 이용한 휴머노이드 로봇의 안정된 보행

정재식 高麗大學校 大學院 2012 국내석사

This paper proposes a stable walking algorithm for humanoid robot with line tracking using reinforcement learning. In the case of walking on the rough surface, gait control is the most important factor for stable walking. There have been many studies about stable walking. Generally, humanoid walking solution is to use inverse kinematics, dynamics, inverted pendulum and ZMP(Zero Moment Point). ZMP uses FSR sensor in most cases but too many data from FSR sensor is not sufficient. The reason is that the response speed of FSR sensor is slow. In this paper, swing value(Shift X) is obtained to adjust data of gyro and acceleration sensor to Q-learning algorithm. We propose method that find optimal path with Q-learning to follow the line.