Seongmin Baek

Motion retargeting and evaluation for VR-based training of free motions

Virtual reality (VR) has emerged as one of the important and effective tools for education and training. Most VR-based training systems are situation based, where the trainees are trained for discrete decision making in special situations presented by the VR environments. In contrast, this paper discusses the application of VR to a different class of training, for learning free motion, often required in sports and the arts. We propose a VR-based motion-training framework that contains an intuitive motion-guiding interface, posture-oriented motion retargeting, and an evaluation and advice scheme for corrective feedback. Applications of the proposed framework to simple fencing training and a dance imitation game are demonstrated .

Implementation of Crowd system in Maya

This paper introduces a practical implementation of virtual crowd scene for contents such as animations, movies, games, and so on. The motion of individual as well as the overall movement of extras is very important because the crowd is gathering of the characters. Animators want to control the overall movement of crowd while each character moves autonomously. Also, they desire to confirm their character models with motions before rendering. However, the characters require lots of memory and consume lots of time because of the model data size, the simulation time, model-import time, motion connection time, etc. Therefore, the crowd system should allow animators to guarantee good motion quality, control the characters easily, and generate crowd scene rapidly. This paper proposes the crowd system that satisfies the conditions

User pose estimation based on multiple depth sensors

Despite the diverse application of motion capture technology, it is challenging to capture peoples motions unless they are wearing relevant equipment. This paper proposes a method of estimating the joint positions based on depth data as well as optimal joint selection in restoring the pose with multiple Kinect sensors. The proposed method enhances the accuracy of pose restoration, enables real-time capture of dynamic motions such as Taekwondo and applies to training programs for the general public.

Real-Time Dynamic Motion Capture Using Multiple Kinects

The present paper proposes a method of capturing real-time motions without any inconvenient suit by using several inexpensive sensors vulnerable to joint occlusion and body rotation. Depth data and ICP algorithm are used for calibration. Then, the left and right sides of joints are determined, and the optimal joints are chosen based on the variation in rotation to restore postures. The similarity between the motions captured by the proposed multiple sensors and those captured by a commercial motion capture system is over 85 %.

Real-time performance capture using multiple Kinects

There have been many cases applying recognized user movements in games as low-cost sensor such as Microsoft Kinect and Asus Xtion have appeared, but there is limitation in capturing user movements with one sensor. This paper uses multiple kinect sensors to propose a method to capture user performance. By solving calibration and synchronization problems between sensors and appropriately combining input data, stable performance capture is possible even in occlusion of body parts.

Golf Ball Bouncing Model Based on Real Images

This paper proposes a more accurate bouncing model based on real trajectories for golf ball. The suggested bouncing model considers the spin of a golf ball, as well as the elasticity and friction of the grass, based on impulse. To measure the accuracy of the model, the trajectories of the ball on a green and fairway were filmed with a camera. As the bouncing model was adjusted based on the real trajectory of the ball extracted from the video, its accuracy is improved.