Mun-Ho Jeong

Human-like catching motion of humanoid using Evolutionary Algorithm(EA)-based imitation learning

A framework to generate a human-like arm motion of a humanoid robot using an Evolutionary Algorithm(EA)-based imitation learning is proposed. The framework consists of two processes, imitation learning of human arm motions and real-time generating of a human-like arm motion using the motion database evolved in the learning process. The imitation learning builds the database for the humanoid robot that is initially converted from human motion capture data and then evolved using a genetic operator based on a Principal Component Analysis (PCA) in an evolutionary algorithm. This evolution process also considers the minimizing of joint torques in the robot. The database is then used to generate humanoid robots arm motions in real-time, which look like humans and require minimal torques. The framework is examined for humanoid robot to reach its arms for catching a ball. Additionally, the inverse kinematics problem to determine the final posture of 6-DOF robot arm with a waist for the task of catching a ball, is proposed.

Improved-Quality Real-Time Stereo Vision Processor

This paper presents a stereo vision processor with the form of ASIC that achieves enhanced quality depth maps and real-time performance. Our vision processor can be used broadly in practical applications. To improve depth map quality, pre- and post-processing units are adopted, and SFRs (Special Function Registers) are assigned to vision parameters for controllable quality. To meet real-time requirements, the stereo vision system is implemented on hardware using sophisticated design. We integrate image rectification, bilateral filtering, depth estimator and left-right consistency check blocks on a single silicon chip. This processor is fabricated in a 0.18-um standard CMOS technology, and can operate at 120MHz clock frequency achieving over 140 frames/s depth maps with 320 by 240 image size and 64 disparity levels. The system exploits 8-bit sub-pixel disparities for depth accuracy, and shows the throughput over 707 million PDS, which is better than results of any published work. The unrectified and unfiltered images taken at real environment are used as test inputs for performance and quality evaluation. Comparisons with previous ASIC implementations are presented to verify the improvement of this task.

Tangible Tele-Meeting in Tangible Space Initiative

Tangible Space Initiative (TSI) is a new framework that can provide a more natural and intuitive Human Computer Interface for users. This is composed of three cooperative components: a Tangible Interface, Responsive Cyber Space, and Tangible Agent. In this paper we present a Tangible Tele-Meeting system in TSI, which allows people to communicate with each other without any spatial limitation. In addition, we introduce a method for registering a Tangible Avatar with a Tangible Agent. The suggested method is based on relative pose estimation between the user and the Tangible Agent. Experimental results show that the user can experience an interaction environment that is more natural and intelligent than that provided by conventional tele-meeting systems.

Architecture and implementation of real-time stereo vision with bilateral background subtraction

We describe the architecture and implementation of bilateral background subtraction for real-time stereo vision system. Pre-smoothing a signal and noise removal may help to improve the performance for many signal-processing algorithms such as compression, detection, enhancement, recognition, and more [2]. Bilateral filtering proposed by C. Tomasi and R. Manduchi can be used as an edge-preserving smoother, removing high-frequency components of an image without blurring its edges [1][3]. Recently, [3] showed enhanced real-time stereo through software implementation of bilateral filtering. In this paper, we show hardware implementation of bilateral background subtraction for real-time stereo and present its architecture as well as required hardware resources. Also, we provide experimental results with real data and present our future works.

A detection cell using multiple points of a rotating triangle to find local planar regions from stereo depth data

This paper presents a method to recognize plane regions for unobstructed motion of mobile robots. When an autonomous agency, using a stereo camera or a laser scanning sensor, is in an unknown 3D environment, the mobile agency must detect the plane regions so that it can independently decide its direction of movement in order to perform assigned tasks. In this paper, a fast method of plane detection is proposed, wherein the normal vector of a triangle is inscribed in a small circular region such that the normal vector passes through the circumcenter area of the triangle. To reduce the effects of noise and outliers, the triangle is rotationally sampled with respect to the center position of the circular region, and a series of inscribed triangles having different normal vectors is generated. The direction vectors of these generated triangles are normalized and the median direction of the normal vectors is then used to test the planarity of the circular region. A pose finding procedure is introduced from range data of a surface to decide the scale and rotation angle of the circular region superimposed on range image data. The method of plane detection is very fast as computation of local information about the plane typically requires sub-ms duration, and the performance of the algorithm for real range data obtained from a stereo camera system has been verified.

Adaptive mean-shift tracking with novel color model

We describe a new method for robust and real time object tracking using a Gaussian cylindroid color model and an adaptive mean shift. Color information has been used for characterizing an object from others. However, sensitiveness to illumination changes limits their flexibility and applicability under various illuminating conditions. We present an effective color space model against irregular illumination changes where chrominance is fitted with respect to intensity using B spline. A target for tracking is expressed by a joint probabilistic density function that incorporates a proposed color space model into the positional space in image lattice. Tracking is performed using the mean shift algorithm where the bandwidth selection is essential to tracking performance. We present a simple and effective method to find the optimal bandwidth that maximizes the lower bound of the log likelihood of the target represented by the joint probabilistic density function. The robustness and capability of the presented method are demonstrated for several image sequences.

Robust Face Tracking Using Bilateral Filtering

This paper introduces a method for face tracking in a video sequence in real time. In this method the profile of color distribution characterizes targets feature. It is invariant for rotation and scale changes. Its also robust to non-rigidity and partial occlusion of the target. We employ the mean-shift algorithm to track the target face and to reduce the computational cost. However, face tracking using color distribution is failed by noises as occlusion including some objects with similar color distribution and with exactly difference color distribution. Thus failures are critical problems. To solve these problems, we employ a bilateral filter which uses the color and range information. We have applied the proposed bilateral filter to track the real time face tracking. The experimental results demonstrate the efficiency of this algorithm. Its performance has been proven superior to the original mean shift tracking algorithm.

A Method for Camera Pose Estimation from Object of a Known Shape

Pose estimation between cameras and object is a central element for computer vision and its applications. In this paper, we present an approach to solve the problem of estimating the camera 3-D location and orientation from a matched set of 3-D model and 2-D image features. We derive an error equation using roll-pitch-yaw angle to present the rotation matrix and directly calculate the partial derivatives of Jacobian matrix without use of numerical methods for estimation parameters from the nonlinear error equation. Because the proposed method does not use a numerical method to derive the partial derivatives, it is very fast and so adequate for real-time pose estimation and also insensitive to selection of initial values for solving the nonlinear equation. The method is proved from real image experiments and a comparison with a numerical estimation method is presented.

Color region tracking against brightness changes

This pater describes a new method for real-time and robust object tracking using a Gaussian-cylindroid color model and an adaptive mean shift. Color information has been widely used for characterizing an object from others. However, sensitiveness to illumination changes limits their flexibility and applicability under various illuminating conditions. We present a robust color model against irregular illumination changes where chrominance is fitted with respect to intensity using B-spline. A target for tracking is expressed by the joint probabilistic density function of the proposed color model and 2-D positional information in image lattice. And tracking is performed using the mean-shift algorithm incorporating the joint probabilistic density function where the bandwidth selection is essential to tracking performance. We present a simple and effective method to find the optimal bandwidth that maximizes the lower bound of the log-likelihood of the target represented by the joint probabilistic density function. The robustness and capability of the presented method are demonstrated for several image sequences.

Gaze from motion: towards natural user interfaces

We propose a method of 3-D gaze estimation allowing the head motion under an uncalibrated monocular camera system. The paper describes the eyeball structure model with compact descriptions of the eyeball motion and its static 3-D structure. Assuming that the eyeball motion is independent of the head motion, we present a dynamic converging-connected model to make the gaze estimation allowing the head motion more systematic and simple. The gaze estimation is performed through the extended Kalman filter using the eyeball structure model and the dynamic converging-connected model. The preliminary test suggests satisfactory results.