Yoo-Joo Choi

SmartBuckle: human activity recognition using a 3-axis accelerometer and a wearable camera

Recognizing human activity is one of the most important concerns in many ubiquitous computing systems. In this paper, we present a wearable intelligence device for medical monitoring applications. We called the SmartBuckle that is designed to recognize human activity and to monitor vitality. We developed human activity recognition algorithms and evaluated them by using data acquired from a 3-axis accelerometer with embedded one image sensor in a belt. In order to evaluate, acceleration data was collected from 9 activity labels. In the image sensor, we extracted activity features based on grid-based optical flow method. In the 3-axis accelerometer sensor, we used the correlation between axes and the magnitude of the FFT for feature extraction. In the experiments, our classifiers showed the excellent performance in recognizing activities with an overall accuracy rate of 93%.

Deciding the Number of Color Histogram Bins for Vehicle Color Recognition

Given vehicle images, we suggest a way to recognize the color of the vehicle contained in the image. The color feature of a vehicle is represented by a color histogram, and we decide the appropriate number of color histogram bins, which mainly affects the successful recognition rate. After generating the histograms, template matching is used to decide the vehicle color. In HSI (hue saturation intensity) color space, experimental results show that the partition of H, S, and I into 8, 4, 4, respectively, achieves the highest success rate up to 88.34%.

Retrieval of Identical Clothing Images Based on Local Color Histograms

In this paper, we present a novel approach to retrieve the person images that contain the identical clothing to a query image from the image set captured by multiple CCTV cameras. In order to measure the similarity of the clothing, we analyze the color information of the clothing area in the image. To find the clothing area, we detect the face area first. The clothing area is found based on the position of the face area. Then, we apply the color quantization to the clothing area. We build six color histograms based on the quantized color for six sub-regions defined in the clothing area. The feature vector for the clothing area is composed by using the color histograms. Similarity between two clothing areas is measured by Euclidean distance between the feature vectors. In the experimental results, our approach shows the better performance compared with the method that uses HSV histogram-based color analysis.

A Robust Hand Recognition In Varying Illumination

As ubiquitous computing provide up-graded smart environments where humans desire to create various types of interaction for many kinds of media and information, the research in the area of Human-Computer Interaction (HCI) is being emphasized to satisfy a more convenient user interface. In particular, the gesture interaction technique has been one of the important research areas under ubiquitous computing environment since it can only utilize widespread consumer video cameras and computer vision techniques without the aid of any other devices to grasp human movements and intentions(Park, 2004; Jung, 2007). Among the gesture interaction techniques, recognition of hand poses and gestures has especially received attention due to great potential to build various and user-centric computer interfaces. The applicability of hand pose recognition is very high in applications where system users can not use existing interface devices such as a keyboard and a mouse since they are required to wear heavy protective gloves for industrial processes. Various types of gesture interfaces have been also presented in three-dimensional games based on virtual reality and these interfaces have enhanced an interest level and creativity within these environments for the users. Humans can distinguish hand poses very quickly through their complex optical systems, while it is very difficult for a computer system to rapidly and accurately understand hand poses. Therefore, many researchers have tried to simulate the human optical system, which can extract objects of interest from complex scenes and understand the context among objects. One of the major factors that disturb automatic gesture recognition is illumination change. The sudden illumination changes lead to the misunderstanding of background and foreground regions. We propose a robust hand recognition technique that can stably extract hand contours even under sudden illumination changes. Figure 1 shows the flowchart for our proposed method. The proposed method acquires the background images for a restricted duration and calculates the mean and standard deviation for the hue and hue-gradient of each pixel within the captured background images. That is, a background model for each pixel is built. The hue and hue-gradient of the input images captured in real-time are calculated and compared to those of the background images. The foreground objects are extracted based on the difference magnitude between those of the input image and the background image. To accurately extract the tight object region of interest, we calculate the eigen value and eigen

Color Correction for Object Identification from Images with Different Color Illumination

This paper presents a novel color correction method for automatic detection of identical objects from images captured in different environments. The proposed method extracts illumination color features using the color board image captured in each illumination and defines conversion vectors from a source illumination to a target illumination in the pre-processing step. In the color correction step, we apply the conversion vectors to RGB elements of each pixel in the object image. In the experiments, we captured two images including a color palette and three target objects in different illumination conditions. We define one image as a source and the other as a target. Illumination conversion vectors for the color correction are defined based on the difference between two color palette images and applied to the source image. In order to evaluate the performance of the proposed color correction, we compared the color differences of identical objects in target, source and corrected images.

Adaptive surface-deformable model with shape-preserving spring

This paper presents a multi‐resolutional surface deformable model with physical property adjustment scheme and shape‐preserving springs to represent surface‐deformable objects efficiently and robustly. In order to reduce the computational complexity while ensuring the same global volumetric behaviour for the deformable object, we introduce a multi‐resolutional mass‐spring model that is locally refined using the modified‐butterfly subdivision scheme. For robust deformation, a shape‐preserving spring, which helps to restore the model to the original shape, is proposed to reduce the animation instability. Volume and shape preservation is indirectly achieved by restoring the model to the original shape without computing the actual volume and associated forces at every iteration. Most existing methods concentrate on visual realism of multi‐resolutional deformation and often neglect to maintain the dynamic behavioural integrity between detail levels. In order to preserve overall physical behaviour, we present a new scheme for adjusting physical properties between different levels of details. During the animation of deformable objects, the part of the object under external forces beyond a threshold or with large surface curvature variations is refined with a higher level of detail. The physical properties of nodes and springs in the locally refined area are adjusted in order to preserve the total mass and global behaviour of the object. The adequacy of the proposed scheme was analysed with tests using practical mesh examples. Experimental results demonstrate improved efficiency in object deformation and preservation of overall behaviour between different levels. Copyright

Cutaneous manifestations of the subtypes of polycystic ovary syndrome in Korean patients

Polycystic ovary syndrome (PCOS) is a common endocrinological disorder in women of childbearing‐age. Although PCOS has common dermatological manifestations, including hirsutism, acne and androgenetic alopecia, little is known about the dermatological characteristics of PCOS patients in Asia.

Adaptive Background Modeling for Effective Ghost Removal and Robust Left Object Detection

A background model using image subtraction in an intelligent video surveillance system could make severe errors in detection and tracking of objects due to changes of natural phenomena such as shadows and wind. Adaptive background models have been proposed in order to solve these problems, but most previous methods can make a ghost and sometimes miss the left objects which have stopped moving for a while. In this paper, we propose an adaptive background method to robustly track left objects and to effectively remove ghosts. The proposed method is based on background subtraction using adaptive median filtering and background update using motion information. In this method, the background is firstly updated based on the motion information in a pixel unit and secondly updated again based on the contour of the objects in a non-motion region unit. The method prevents the left objects from absorbing into the background and removes the ghosts quickly. In the experiments, we prove an effectiveness of our method through the comparison with the previous adaptive median filtering background subtraction.

Rapid pairwise intersection tests using programmable GPUs

Detecting self-intersections within a triangular mesh model is fundamentally a quadratic problem in terms of its computational complexity, since in principle all triangles must be compared with all others. We reflect the 2D nature of this process by storing the triangles as multiple 1D textures in texture memory, and then exploit the massive parallelism of graphics processing units (GPUs) to perform pairwise comparisons, using a pixel shader. This approach avoids the creation and maintenance of auxiliary geometric structures, such as a bounding volume hierarchy (BVH); but nevertheless we can plug in auxiliary culling schemes, and use stencils to indicate triangle pairs that do not need to be compared. To overcome the readback bottleneck between GPU and CPU, we use a hierarchical encoding scheme. We have applied our technique to detecting self-intersections in extensively deformed models, and we achieve an order of magnitude increase in performance over CPU-based techniques such as [17].

self -CD: interactive self-collision detection for deformable body simulation using GPUs

This paper presents an efficient self-collision detection algorithm for deformable body simulation using programmable graphics processing units(GPUs). The proposed approach stores a triangular mesh representation of a deformable model as 1D textures and rapidly detects self-collisions between all pairs of triangular primitives using the programmable SIMD capability of GPUs [1]. Since pre-computed spatial structure such as bounding volume hierarchy is not used in our algorithm, our algorithm does not require expensive runtime updates to such complex structure as the underlying model deforms. Moreover, in order to overcome a potential bottleneck between CPU and GPU, we propose a hierarchical encoding/decoding scheme using multiple off-screen buffers and multi-pass rendering techniques, which reads only a region of interests in the resulting off-screen buffer.