1. Detection of Calibration Pattern using Concentric Circles and Bimodal Distribution of Intensity Histogram
A calibration pattern detection method is proposed wherein a bimodal distribution is approximated for the intensity histogram of the calibration pattern. Because the histogram of the calibration pattern, such as black and white squares, is assumed to be bimodal, pattern detection is employed to determine the best threshold separating the two modes of the histogram. As most real images may include uneven illumination, the method divides the original image into subimages and then utilizes a different threshold to segment each subimage and approximate image parts of well separated modes. For binary images obtained from different threshold values of different subimages, region labeling and boundary tracing of labeled regions are performed and ellipses with close centers are detected as reference points of the calibration pattern among many background clutters.
2. 초음파 무선 센서노드를 이용한 실시간 위치 추적 시스템
Location information will become increasingly important for future Pervasive Computing applications. Location tracking system of a moving device can be classified into two types of architectures: an active mobile architecture and a passive mobile architecture. In the former, a mobile device actively transmits signals for estimating distances to listeners. In the latter, a mobile device listens signals from beacons passively. Although the passive architecture such as Cricket location system is inexpensive, easy to set up, and safe, it is less precise than the active one. In this paper, we present a passive location system using Cricket Mote sensors which use RF and ultrasonic signals to estimate distances. In order to improve accuracy of the passive system, the transmission speed of ultrasound was compensated according to air temperature at the moment. Upper and lower bounds of a distance estimation were set up through measuring minimum and maximum distances that ultrasonic signal can reach to. Distance estimations beyond the upper and the lower bounds were filtered off as errors in our scheme. With collecting distance estimation data at various locations and comparing each distance estimation with real distance respectively, we proposed an equation to compensate the deviation at each point. Equations for proposed algorithm were derived to calculate relative coordinates of a moving device. At indoor and outdoor tests, average location error and average location tracking period were 3.5 cm and 0.5 second, respectively, which outperformed Cricket location system of MIT.
3. Abnormal behavior detection using Gaussian Mixture Model and Optical Flow
본 논문에서는 감시시스템이 갖추어진 환경 내에서 발생할 수 있는 특이 행동을 효율적으로 감지하기 위한 기법을 제시한다. 최근 대형 범죄 및 방화 사건 등의 방지목적으로 DVR 의 단순 녹화를 벗어나 지능형 감시시스템을 도입하려는 연구가 활발히 진행되고 있다. 그러나 이러한 시스템들은 아직 초기 연구 단계에 있으며 영상내의 관심물체 추출을 위한 전경과 배경의 분리 및 추적 단계에 그치고 있다. 이에 본 논문에서는 가우시안 혼합 모델을 통하여 전경과 배경을 분리하고, 관심영역에 한해서 Optical Flow 기법을 이용하여 폭력상황과 같은 특이 행동의 감지 여부를 판단 할 수 있는 방법에 대해 실험을 통해 평가하였다.
4. METHOD TO IMPROVE THE PERFORMANCE OF THE ADABOOST
ALGORITHM USING GAUSSIAN PROBABILITY DISTRIBUTION
The weak classifier of AdaBoost algorithm is a central classification element that uses a single criterion separating positive and negative learning candidates. Finding the best criterion to separate two feature distributions influences learning capacity of the algorithm. A common way to classify the distributions is to use the mean value of the features. However, positive and negative distributions of Haar-like feature as an image descriptor are hard to classify by a single threshold. The poor classification ability of the single threshold also increases the number of
boosting operations, and finally results in a poor classifier. This paper propose the probability AdaBoost Algorithm that is made the Gaussian probability distribution of feature value and evaluate the probability value as how to close the
mean of the Gaussian probability distribution. In the learning procedure, the weak classifier is selected by the evaluation that is how positive distribution to become independent negative distribution and how positive distribution to close. The weight is updated to exponential ‘0’ or ‘1’ in conventional AdaBoost but the proposal method is updated to exponential the real value between ‘0’ and ‘1’ by the Gaussian distribution. Hence, the selection of weak classifier is reflected more preciously to weight update. It is no specific threshold to study the proposed method using the Gaussian
probability distribution of positive feature value. It is learned by 2 distribution of positive and negative date; therefore, The modeling for to classify the positive is more natural. and we prove more previously detection in experiment.
5. Camera Calibration Method under Poor Lighting Condition in Factories
This paper proposes a method to perform accurate camera calibration under poor lighting condition of factories or industrial fields. Preprocessing of camera calibration required for measuring object dimensions has to be able to extract calibration points from patterns of the calibration scale, for example, the calibration from plane pattern scale needs at least seven points of the known dimension marked on the scale. However, industrial fields hardly provide proper lighting condition for camera calibration of the measurement system. The data points for calibration are
automatically selected from a probabilistic assumption for size variation of the calibration point when the threshold level changes for image binarization. The system requires user to provide at least four points that are incomplete, these points are used to predict position of exact calibration points and extract accurate calibration parameters in an iterative procedure using nonlinear optimization of the parameters. From real images, we prove the method can be applied to camera calibration of poor quality images obtained under lens distortion and bad illumination.
6. A Method to Detect Multiple Plane Areas by using the Iterative Randomized Hough Transform(IRHT) and the Plane Detecton.