Open Science Research Excellence

Keechul Jung

Publications

8

Publications

8
4156
Real-time 3D Feature Extraction without Explicit 3D Object Reconstruction
Abstract:
For the communication between human and computer in an interactive computing environment, the gesture recognition is studied vigorously. Therefore, a lot of studies have proposed efficient methods about the recognition algorithm using 2D camera captured images. However, there is a limitation to these methods, such as the extracted features cannot fully represent the object in real world. Although many studies used 3D features instead of 2D features for more accurate gesture recognition, the problem, such as the processing time to generate 3D objects, is still unsolved in related researches. Therefore we propose a method to extract the 3D features combined with the 3D object reconstruction. This method uses the modified GPU-based visual hull generation algorithm which disables unnecessary processes, such as the texture calculation to generate three kinds of 3D projection maps as the 3D feature: a nearest boundary, a farthest boundary, and a thickness of the object projected on the base-plane. In the section of experimental results, we present results of proposed method on eight human postures: T shape, both hands up, right hand up, left hand up, hands front, stand, sit and bend, and compare the computational time of the proposed method with that of the previous methods.
Keywords:
Fast 3D Feature Extraction, Gesture Recognition,Computer Vision.
7
4451
3D Star Skeleton for Fast Human Posture Representation
Abstract:
In this paper, we propose an improved 3D star skeleton technique, which is a suitable skeletonization for human posture representation and reflects the 3D information of human posture. Moreover, the proposed technique is simple and then can be performed in real-time. The existing skeleton construction techniques, such as distance transformation, Voronoi diagram, and thinning, focus on the precision of skeleton information. Therefore, those techniques are not applicable to real-time posture recognition since they are computationally expensive and highly susceptible to noise of boundary. Although a 2D star skeleton was proposed to complement these problems, it also has some limitations to describe the 3D information of the posture. To represent human posture effectively, the constructed skeleton should consider the 3D information of posture. The proposed 3D star skeleton contains 3D data of human, and focuses on human action and posture recognition. Our 3D star skeleton uses the 8 projection maps which have 2D silhouette information and depth data of human surface. And the extremal points can be extracted as the features of 3D star skeleton, without searching whole boundary of object. Therefore, on execution time, our 3D star skeleton is faster than the “greedy" 3D star skeleton using the whole boundary points on the surface. Moreover, our method can offer more accurate skeleton of posture than the existing star skeleton since the 3D data for the object is concerned. Additionally, we make a codebook, a collection of representative 3D star skeletons about 7 postures, to recognize what posture of constructed skeleton is.
Keywords:
computer vision, gesture recognition, skeletonization,human posture representation.
6
4486
Real-Time Vision-based Korean Finger Spelling Recognition System
Abstract:
Finger spelling is an art of communicating by signs made with fingers, and has been introduced into sign language to serve as a bridge between the sign language and the verbal language. Previous approaches to finger spelling recognition are classified into two categories: glove-based and vision-based approaches. The glove-based approach is simpler and more accurate recognizing work of hand posture than vision-based, yet the interfaces require the user to wear a cumbersome and carry a load of cables that connected the device to a computer. In contrast, the vision-based approaches provide an attractive alternative to the cumbersome interface, and promise more natural and unobtrusive human-computer interaction. The vision-based approaches generally consist of two steps: hand extraction and recognition, and two steps are processed independently. This paper proposes real-time vision-based Korean finger spelling recognition system by integrating hand extraction into recognition. First, we tentatively detect a hand region using CAMShift algorithm. Then fill factor and aspect ratio estimated by width and height estimated by CAMShift are used to choose candidate from database, which can reduce the number of matching in recognition step. To recognize the finger spelling, we use DTW(dynamic time warping) based on modified chain codes, to be robust to scale and orientation variations. In this procedure, since accurate hand regions, without holes and noises, should be extracted to improve the precision, we use graph cuts algorithm that globally minimize the energy function elegantly expressed by Markov random fields (MRFs). In the experiments, the computational times are less than 130ms, and the times are not related to the number of templates of finger spellings in database, as candidate templates are selected in extraction step.
Keywords:
CAMShift, DTW, Graph Cuts, MRF.
5
4503
Fast 3D Collision Detection Algorithm using 2D Intersection Area
Abstract:

There are many researches to detect collision between real object and virtual object in 3D space. In general, these techniques are need to huge computing power. So, many research and study are constructed by using cloud computing, network computing, and distribute computing. As a reason of these, this paper proposed a novel fast 3D collision detection algorithm between real and virtual object using 2D intersection area. Proposed algorithm uses 4 multiple cameras and coarse-and-fine method to improve accuracy and speed performance of collision detection. In the coarse step, this system examines the intersection area between real and virtual object silhouettes from all camera views. The result of this step is the index of virtual sensors which has a possibility of collision in 3D space. To decide collision accurately, at the fine step, this system examines the collision detection in 3D space by using the visual hull algorithm. Performance of the algorithm is verified by comparing with existing algorithm. We believe proposed algorithm help many other research, study and application fields such as HCI, augmented reality, intelligent space, and so on.

Keywords:
Collision Detection, Computer Vision, Human Computer Interaction, Visual Hull
4
4535
View-Point Insensitive Human Pose Recognition using Neural Network
Abstract:
This paper proposes view-point insensitive human pose recognition system using neural network. Recognition system consists of silhouette image capturing module, data driven database, and neural network. The advantages of our system are first, it is possible to capture multiple view-point silhouette images of 3D human model automatically. This automatic capture module is helpful to reduce time consuming task of database construction. Second, we develop huge feature database to offer view-point insensitivity at pose recognition. Third, we use neural network to recognize human pose from multiple-view because every pose from each model have similar feature patterns, even though each model has different appearance and view-point. To construct database, we need to create 3D human model using 3D manipulate tools. Contour shape is used to convert silhouette image to feature vector of 12 degree. This extraction task is processed semi-automatically, which benefits in that capturing images and converting to silhouette images from the real capturing environment is needless. We demonstrate the effectiveness of our approach with experiments on virtual environment.
Keywords:
Computer vision, neural network, pose recognition,view-point insensitive.
3
5347
View-Point Insensitive Human Pose Recognition using Neural Network and CUDA
Abstract:
Although lots of research work has been done for human pose recognition, the view-point of cameras is still critical problem of overall recognition system. In this paper, view-point insensitive human pose recognition is proposed. The aims of the proposed system are view-point insensitivity and real-time processing. Recognition system consists of feature extraction module, neural network and real-time feed forward calculation. First, histogram-based method is used to extract feature from silhouette image and it is suitable for represent the shape of human pose. To reduce the dimension of feature vector, Principle Component Analysis(PCA) is used. Second, real-time processing is implemented by using Compute Unified Device Architecture(CUDA) and this architecture improves the speed of feed-forward calculation of neural network. We demonstrate the effectiveness of our approach with experiments on real environment.
Keywords:
computer vision, neural network, pose recognition,view-point insensitive, PCA, CUDA.
2
12456
Human Pose Estimation using Active Shape Models
Abstract:

Human pose estimation can be executed using Active Shape Models. The existing techniques for applying to human-body research using Active Shape Models, such as human detection, primarily take the form of silhouette of human body. This technique is not able to estimate accurately for human pose to concern two arms and legs, as the silhouette of human body represents the shape as out of round. To solve this problem, we applied the human body model as stick-figure, “skeleton". The skeleton model of human body can give consideration to various shapes of human pose. To obtain effective estimation result, we applied background subtraction and deformed matching algorithm of primary Active Shape Models in the fitting process. The images which were used to make the model were 600 human bodies, and the model has 17 landmark points which indicate body junction and key features of human pose. The maximum iteration for the fitting process was 30 times and the execution time was less than .03 sec.

Keywords:
Active shape models, skeleton, pose estimation.
1
15583
Graph-based High Level Motion Segmentation using Normalized Cuts
Abstract:
Motion capture devices have been utilized in producing several contents, such as movies and video games. However, since motion capture devices are expensive and inconvenient to use, motions segmented from captured data was recycled and synthesized to utilize it in another contents, but the motions were generally segmented by contents producers in manual. Therefore, automatic motion segmentation is recently getting a lot of attentions. Previous approaches are divided into on-line and off-line, where on-line approaches segment motions based on similarities between neighboring frames and off-line approaches segment motions by capturing the global characteristics in feature space. In this paper, we propose a graph-based high-level motion segmentation method. Since high-level motions consist of several repeated frames within temporal distances, we consider all similarities among all frames within the temporal distance. This is achieved by constructing a graph, where each vertex represents a frame and the edges between the frames are weighted by their similarity. Then, normalized cuts algorithm is used to partition the constructed graph into several sub-graphs by globally finding minimum cuts. In the experiments, the results using the proposed method showed better performance than PCA-based method in on-line and GMM-based method in off-line, as the proposed method globally segment motions from the graph constructed based similarities between neighboring frames as well as similarities among all frames within temporal distances.
Keywords:
Capture Devices, High-Level Motion, Motion Segmentation, Normalized Cuts