Youngsung Soh

Multi-Target Tracking Using Color Information

—For security purposes, it is prerequisite to track multiple targets efficiently. Most of the current implementation uses Kalman filter and color information independently. The proposed method combines extended Kalman filter and color information for tracking multiple objects under high occlusion. For tracking, the first thing done is the object detection. The background model used to segment foreground from background is spatio-temporal Gaussian mixture model (STGMM). Tracking consists of two steps: independent object tracking and occluded object tracking. For independent object tracking we exploit extended Kalman filter, whereas for occluded object tracking, color information attribute is used. The system was tested in real world application and successful results were obtained.

Spatio-temporal Gaussian Mixture Model for Background Modeling

Background subtraction is widely employed in the detection of moving objects when background does not show much dynamic behavior. Many background models have been proposed by researchers. Most of them analyses only temporal behavior of pixels and ignores spatial relations of neighborhood that may be a key to better separation of foreground from background when background has dynamic activities. To remedy, some researchers proposed spatio-temporal approaches usually in the block-based framework. Two recent reviews[1, 2] showed that temporal kernel density estimation(KDE) method and temporal Gaussian mixture model(GMM) perform about equally best among possible temporal background models. Spatio-temporal version of KDE was proposed. However, for GMM, explicit extension to spatio-temporal domain is not easily seen in the literature. In this paper, we propose an extension of GMM from temporal domain to spatio-temporal domain. We applied the methods to well known test sequences and found that the proposed outperforms the temporal GMM.

Tracking Occluded Objects Using Kalman Filter and Color Information

—Robust visual tracking is imperative to track multiple occluded objects. Kalman filter and color information tracking algorithms are implemented independently in most of the current research. The proposed method combines extended Kalman filter with past and color information for tracking multiple objects under high occlusion. The proposed method is robust to background modeling technique. Object detection is done using spatio-temporal Gaussian mixture model (STGMM). Tracking consists of two steps: partially occluded object tracking and highly occluded object tracking. Tracking partially occluded objects, extended Kalman filter is exploited with past information of object, whereas for highly occluded object tracking, color information and size attributes are used. The system was tested in real world application and successful results were obtained.

Development of an Enhanced Obstacle Avoidance Algorithm for a Network-Based Autonomous Mobile Robot

An enhanced obstacle avoidance algorithm for a network-based autonomous mobile robot is proposed in this paper. Firstly, the readings of the environmental sensors at a moment are compensated to the prospecting readings of the sensors considering network delay measured and the kinematic model of the robot. The compensated readings of the sensors are used for building the polar histogram of the VFH algorithm. Secondly, a sensory fusion using the Kalman filter is proposed for the localization of the robot to compensate both the delay of the readings of an odometry sensor and the delay of the readings of the environmental sensors. The performance enhancements of the proposed obstacle avoidance algorithm from the viewpoint of efficient path generation and accurate goal positioning are also shown in this paper through some simulation tests.

A Featured Area-Based Image Registration

Abstract—Image registration is necessary when images from multiple viewpoints need be brought into common coordinate system. Image registration techniques can be classified as area-based methods and feature-based methods. In area-based methods, no features are selected and regularly tessellated areas are usually used for matching. In feature-based methods, features such as regions, lines, and prominent points are detected and used for matching. When image contains rich features, feature-based methods are preferred and when it does not, area-based methods are usually adopted. There are occasions where richness of features varies locally in the image. In this case, either area-based methods or feature-based methods alone may not generate successful results. In this paper, we propose a mixture of two methods termed as featured area-based method. In the proposed, we first tessellate the image into equal-sized areas, estimate richness of features of each area utilizing the edge direction histogram, choose only those areas with a certain level of richness, and use them for matching. We compared the proposed with well-known conventional methods such as Kanade-Lucas-Tomasi(KLT) method, speeded up robust features (SURF), and scale-invariant feature transform (SIFT), and showed that the proposed performs better than others.

A Programmable Multi-Format Video Decoder

In this paper, we propose a VLSI design of Programmable Multi-Format Video Decoder (PMD) to support video codec standards such as MPEG-2, MPEG-4,H.264, and VC-1. It is a hardware and software hybrid system by moving large portion of the complicated data paths, control logics, and computations into processors as software. More flexibility and expandability of the design is achieved by increasing software portion. The required performance for Full-HD video decoding is delivered by the assistance of hardwired logic blocks. The burden of hardwired logic design is reduced.

A New Parallel Implementation of DSI Based Disparity Computation Using CUDA

Stereo matching techniques are used to extract 3D information from 2D stereo pair of images. It can be classified into feature based approach, window (area) based approach, and optimization based approach. Feature based approach generally generates sparse disparity map with high accuracy and low execution time. Window based approach produces dense disparity map with low accuracy and low execution time. Optimization based approach generates dense disparity map with high accuracy and high execution time. Since the ultimate goal of stereo matching is to obtain dense disparity map with high accuracy and low execution time, we choose to select optimization based approach and implement it in parallel framework to overcome execution speed deficiency. There are several optimization methods including dynamic programming, energy minimization, and graph algorithms. We choose to use dynamic programming based on disparity space image (DSI) since it is most appropriate for parallel framework. In this paper, we propose a new parallel algorithm and framework for DSI construction, dynamic programming (DP), and disparity computation using Compute Unified Device Architecture (CUDA). We tested the method on several stereo pairs and found that the method shows remarkable speedup while preserving the quality at a reasonable level.

A Hybridized Disparity Computation Fusing Disparity Space Image and Multi Resolution Image Segmentation

D information is getting more importance nowadays.There are several ways to get 3D information.One way to get 3D information involves the laser based depth estimators, which is a very costly process. Other way is to extract 3D information from 2D stereo image pairs by using stereo matching techniques. Stereo matching is a vast area of research. It can be classified into area (window) based approach, feature based approach, and optimization based approach. Area based approach generally generates dense disparity map with low accuracy and low computation time. Feature based method produces highly accurate and sparse disparity map with low computation time.Optimization based method produces dense disparity map with high accuracy and high execution time.By keeping these things in mind, we proposed a new hybrid way for disparity computation. The proposed method consists of three steps. Theyare disparityestimation, image segmentation, and disparity refinement. Since the ultimate goal of stereo matching is to obtain dense disparity map with high accuracy and low execution time, we select optimization based approach for disparity estimation step and for image segmentation step we select multi resolution image segmentation. At the end, disparity refinement is done by combining the result of both the previous steps.As there are several optimization techniques, we choose disparity space image (DSI) baseddynamic programming (DP).We tested the proposed method on several stereo pairs and found that method produced reasonably good quality results. Index Terms—Stereo matching, DSI, DP.

Gender Classification Using Faces and Gaits

Gender classification is one of the challenging problems in computer vision. Many interactive applications need to exactly recognize human genders. In this paper, we are carrying out some experiments to classify the human gender in conditions of low captured video resolution. We use Local Binary Pattern, Gray Level Co-occurrence Matrix to extract the features from faces and Gait Energy Motion, Gait Energy Image for gaits. We propose to combine face and gait features with the combination classifier to enhance gender classification performance.