International Journal on Artificial Intelligence Tools

In this paper a person-specific saliency system and subsequently two architectures for the recognition of dynamic gestures are described. The systems implemented are designed to take a sequence of images and to assign it to one of a number of discrete classes where each of them corresponds to a gesture from a predefined small vocabulary. Since we think that for a human-computer interaction the localization of the user is essential for any further step regarding the recognition and the interpretation of gestures, in the first part, we begin with describing our saliency system dedicated to the person localization task in cluttered environments. Successively, the intrinsic gesture recognition process is broken down into an initial preprocessing stage followed by a mapping from the preprocessed input variables to an output variable representing the class label. Subsequently, we utilize two different classifiers for mapping the ordered sequence of feature vectors to one gesture category. The first classifier utilizes a hybrid combination of Kohonen Self-Organizing Map (SOM) and Discrete Hidden Markov Models (DHMM). As second recognizer a system of Continuous Hidden Markov Models (CHMM) is used. Preliminary experiments with our baseline systems are demonstrated.

In recent years because of the advances in computer vision research, free hand gestures have been explored as a means of human-computer interaction (HCI). Gestures in combination with speech can be an important step toward natural, multimodal HCI. However, interpretation of gestures in a multimodal setting can be a particularly challenging problem. In this paper, we propose an approach for studying multimodal HCI in the context of a computerized map. An implemented testbed allows us to conduct user studies and address issues toward understanding of hand gestures in a multimodal computer interface.

Absence of an adequate gesture classification in HCI makes gesture interpretation difficult. We formalize a method for bootstrapping the interpretation process by a semantic classification of gesture primitives in HCI context. We distinguish two main categories of gesture classes based on their spatio-temporal deixis. Results of user studies revealed that gesture primitives, originally extracted from weather map narration, form patterns of co-occurrence with speech parts in association with their meaning in a visual display control system. The results of these studies indicated two levels of gesture meaning: individual stroke and motion complex. These findings define a direction in approaching interpretation in natural gesture-speech interfaces.

We consider the problem of searching for the face of a particular individual in a two-dimensional intensity image. This problem has many potential applications such as locating a person in a crowd using images obtained by surveillance cameras. There are two steps in solving this problem: first, face regions must be extracted from the image(s) (face detection) and second, candidate faces must be compared against a face of interest (face verification). Without any a-priori knowledge about the location and size of a face in an image, every possible image location and face size should be considered, leading to a very large search space. In this paper, we propose using Genetic Algorithms (GAs) for searching the image efficiently. Specifically, we use GAs to find image sub-windows that contain faces and in particular, the face of interest. Each sub-window is evaluated using a fitness function containing two terms: the first term favors sub-windows containing faces while the second term favors sub-windows containing faces similar to the face of interest. Both terms have been derived using the theory of eigenspaces. A set of increasingly complex scenes demonstrate the performance of the proposed genetic-search approach.

We propose a method that aids in the evaluation of residual strain in blood vessels. Strain measurement is important in stress analysis of cardiovascular tissues. Since its introduction in 1983, arterial residual strain is evaluated manually which is a lengthy process. This involved photography, film development, printing of large photographs, and finally digitization of the edges of the vascular rings before and after cut using a flatbed digitizing tablet. These processes are highly labor intensive and time consuming, in the order of days. In this paper we introduce an image analysis and processing technique in which arterial residual strain can be evaluated efficiently and reliably. To insure portability, a JAVA-based application is developed which is designed for a semi-automatic operation in evaluating residual strain values. The use of the image-based evaluation technique is validated by comparison of arterial residual strain evaluated using photographs and digitizing tablet.

An algorithm of cancer recognition of ultrasound images is developed in this paper. As we pointed out in our previous work, in order for cancer to survive it develops its own blood supply system, which is different than the supply system of normal tissue. The velocity of the blood flowing through the cancerous blood vessels is different than the velocity of the blood flowing through blood vessels of normal tissue. Due to this fact the ultrasound signal is absorbed differently in the cancerous areas than in the normal tissue areas. The energy of the signal, the continuity of the signal, the autocorrelation function and frequency domain properties are different in the normal tissue than in cancerous tissue. All of these indicators are weighted here for the purpose of classifying the image of the tissue as being cancerous or non cancerous. Preliminary results based on limited number of ultrasound images show that our method has the ability to recognize cancer in ultrasound images.

This paper presents an efficient technique for unsupervised content-based segmentation in stereoscopic video sequences by appropriately combined different content descriptors in a hierarchical framework. Three main modules are involved in the proposed scheme; extraction of reliable depth information, image partition into color and depth regions and a constrained fusion algorithm of color segments using information derived from the depth map. In the first module, each stereo pair is analyzed and the disparity field and depth map are estimated. Occlusion detection and compensation are also applied for improving the depth map estimation. In the following phase, color and depth regions are created using a novel complexity-reducing multiresolution implementation of the Recursive Shortest Spanning Tree algorithm (M-RSST). While depth segments provide a coarse representation of the image content, color regions describe very accurately object boundaries. For this reason, in the final phase, a new segmentation fusion algorithm is employed which projects color segments onto depth segments. Experimental results are presented which exhibit the efficiency of the proposed scheme as content-based descriptor, even in case of images with complicated visual content.

We present a multi-agent system architecture dedicated to model computer vision systems, which provides the vision system with a great degree of flexibility. The basic idea of this architecture is to model a vision system as a society of autonomous agents, where each agent is responsible for specific vision tasks, the control strategy of a vision system is decentralized, and agents communicate using a flexible but easy understandable communication language. This directly leads to self-organizing vision systems, which accomplish vision tasks by goal-driven communication processes. We describe in detail the basic concepts of the proposed multi-agent system approach including the agent architectures, the communication language and network, as well as the interaction strategies. As a testbed for the proposed architecture we have modeled an object recognition system as an assembly of agents which organize themselves according to a given recognition task by employing communication.