Hyoungseop Kim

Human activity recognition: Various paradigms

Action and activity representation and recognition are very demanding research area in computer vision and man-machine interaction. Though plenty of researches have been done in this arena, the field is still immature. Over the last decades, extensive research methodologies have been developed on human activity analysis and recognition for various applications. This paper overviews various recent methods for human activity recognition with analysis. We attempt to sum up the various methods related to human motion representation and recognition. We make an effort to categorize the recent methods from the best in the business, and finally figure out the short-comings and challenges to dig out in future to develop robust action recognition approaches. This work exclusively endeavors to encompass the researches related only to human action recognition mainly from 2001 till-to-date with critical assessment of the methods. We also present our work along with to solve some of the shortcomings. It will widely benefit the researchers to understand and compare the related advancements in this area.

A moving object tracking based on color information employing a particle filter algorithm

In this article, we present a new algorithm to track a moving object based on color information employing a particle filter algorithm. Recently, a particle filter has been proven very successful for nonlinear and non-Gaussian estimation problems. It approximates a posterior probability density of the state, such as the object position, by using samples which are called particles. The probability distribution of the state of the tracked object is approximated by a set of particles, where each state is denoted as the hypothetical state of the tracked object and its weight. The particles are propagated according to a state space model. Here, the state is treated as the position of the object. The weight is considered as the likelihood of each particle. For this likelihood, we consider the similarity between the color histogram of the tracked object and the region around the position of each particle. The Bhattacharya distance is used to measure this similarity. Finally, the mean state of the particles is treated as the estimated position of the object. Experiments were performed to confirm the effectiveness of this method to track a moving object.

Directional motion history templates for low resolution motion recognition

Human motion recognition in low-resolution video is very difficult task because due to low-resolution we miss much significant motion information. In this paper, we demonstrated appearance-based directional motion history image (DMHI) method to recognize various levels of video resolutions. The DMHI technique can overcome the self-occlusion problem that arises from motion overwriting. We found that it can significantly solve motion overwriting problem and can recognize complex actions or activities. We employed the same datasets with various levels of resolutions to test the DMHI and achieved satisfactory result up to a limit. When the resolution is very low, due to significant loss in motion information, we came across some difficulties to recognize the actions properly.

Action recognition with various speeds and timed-DMHI feature vectors

Usually, various motion recognition approaches can not perform well on actions that have variable speeds, or part of the dataset is speedy and vice versa. In this paper, we present the timing issue of our Directional Motion History Image method. This method can solve overwriting or motion self-occlusion problem significantly and thereby it performs well for complex and repetitive activities. However, it is important to analyze the method with activities having various speeds to show its robustness. The experimental results demonstrate that it can perform well with variable paces of the actions though the recognition rate is compromised a bit compared to the dataset having usual speed. We also improve the classification method with the incorporation of motion duration in the final feature vector so that for similar type of activities having different pace, it can show better result. Experiments on another dataset show better performance with the timing incorporation. The achieved recognition rates are very encouraging for further research and implementation is some intelligent systems for surveillance and related applications.

Solutions to motion self-occlusion problem in human activity analysis

Human motion self-occlusion due to motion overlapping in the same region is a daunting task to solve. Various motion-recognition methods either bypass this problem or solve this problem in complex manner. Appearance-based template matching paradigms are simpler and hence faster approaches for activity analysis. In this paper, we concentrate on motion self-occlusion problem due to motion overlapping in various complex activities for recognition. This paper illustrates the directional motion history image concept and compares this motion representation approach with multilevel motion history representation and hierarchical motion history histogram representation to solve the self-occlusion problem of basic motion history image representation. We employ some complex aerobics and find the robustness of our method compared to other methods for this self-occlusion problem. We employ seven higher order Hu moments to compute the feature vector for each activity. Afterwards, k-nearest neighbor method is utilized for classification with leave-one-out paradigm. The comparative results clearly demonstrate the superiority of our method than other recent approaches.

Three-dimensional human motion modeling by back-projection based on image-based camera calibration

This article proposes a back-projection technique for the modeling of 3-D human motion that performs camera calibration using the multiple video sequences obtained. This technique calculates an affine camera matrix by the factorization method, and performs back-projection under the affine camera model. The proposed technique needs neither a 3-D camera calibration tool nor markers for shape recovery, and can recover human motion from silhouette images. We also propose a shadow detector and eliminator using color information and normalized cross correlation for the robust extraction and elimination of shadows. Experimental results show the effectiveness of the proposed technique.

Detection of blood vessels on CTA images of the legs

The disease that causes the obstruction of blood flow by arteriosclerosis and thickening of the arteries in the legs is called arteriosclerosis obliterans (ASO). The early detection and treatment of ASO is very important issue in the medical field. Recently, by using computer aided diagnosis (CAD), physicians can easily detect the blood vessel on the image and displaying the results by use of the system. One of the techniques for analyzing the blood vessels widely used in medical imaging is computed tomography angiography (CTA). The CTA uses three dimensional (3-D) imaging technologies. It also produces a clear image of main blood vessels throughout the body by using contrast media. The 3-D blood vessel image is made the volume data obtained by the CTA. By using 3-D image, we can observe the image from multi aspect. And also, it leads to the improvement of the diagnosis accuracy. In order to diagnose a symptom of the ASO, we propose a method for visual screening technique by detecting blood vessel area on the images. In this paper, we have developed a technique for separation of bone region and blood vessel area in order to extract blood vessels from a CTA volume image. The proposed technique was applied to two real CTA cases and satisfactory results for segmentation of the blood vessels were obtained. Some experimental results are shown with discussions.

Extraction of multi organs by use of level set method from CT images

Recently, various imaging equipments have been introduced into medical fields. Especially, HRCT is one of the most useful diagnosis systems because it provides a high resolution image to physicians. Accordingly, many related image processing techniques are proposed into medical fields for extraction of abnormal area. In the medical image processing field, segmentation is one of the most important problems for analyzing the abnormalities and recognition of internal structures before the operation. Many related segmentation techniques have been developed for automatic extraction of regions of interest. Especially, in order to extract multi organs and to understand the structure of them, several approaches have been developed in the past. But there are still no fully automatic segmentation methods that are generally applicable to regions of interest based on CT image set. In this paper, we propose a new technique for automatic extraction of the multi organs on the MDCT images employing the level set method. We apply the proposed technique to three CT cases and satisfactory results are achieved.

Development of the MI-Viewer KIT for medical image viewer

Recently, the performance of CT (Computed Tomography) scanner with high resolution is rapidly introduced to medical field for detecting the abnormalities. This improvement can make the diagnosis more accurately. To diagnosis the human body, radiologist compares the anatomical and functional image by use of image viewer system in the medical field. The CT scanner is able to produce huge numbers of CT images at once on the visual screening. Therefore, it takes time for to diagnose a patient by using the CT images directly. To overcome this problem, the CAD (computer aided diagnosis) system is developed and admitted effectively as a second opinion for physicians. However, this CAD system is too expensive for introducing to a small hospital user and also needs a computer with high specification for 3-D displaying of CT image sets. To avoid this problem, we develop a new user-friendly CAD system. Our proposed CAD system can be implemented in a personal computer. In this paper, we introduce the CAD system, such as, segmentation by using a Snakes and 3-D display by using a volume rendering technique. The segmentation process enables us to segment the image by connecting the image contour using initial point setting semi-automatically. The 3-D image which is obtained by proposed technique can be supported to realize internal structure of the human body. By using our proposed system, radiologist can be segment multi organs and analyzing the internal human structures.

Performance analysis on an efficient human motion database with various motion representations

In this paper, our proposed structured human motion database is adopted for different motion representations. The motions are first represented as a sequence of frames of 2D images, which were compressed using three recognized motion representation techniques: exclusive-OR, MEI (motion energy image), and MHI (motion history images). The representation is a 2D feature image. The feature image is compressed by characterizing the eigenvectors. A complete vector space called an eigenspace is constructed that represents the image feature vectors for the feature image. The motions are indexed using the projections onto the eigenspace. For the purpose of efficient searching within the database, our proposed B-tree motion database is created and maintained. The comparative performance evaluations for the aforesaid representations were investigated and satisfactory performances (about 90% recognition rate and smaller searching time) were realized for all of the cases using our proposed motion database structure.