COMPUTER VISION
examines and acts upon images with a computer, i.e., its applications require
computers to use the visual information directly without involving any
human being in the visual loop. Computer vision is no longer a rather small
and exclusive research area. Motion tracking and Image analysis
are two of the major topics within this field and are the topics within
which the
CVIP
is working on. Fast processing response is a major requirement in many
computer vision applications. Typical vision systems involve real-time
processing where a sequence of image frames must be processes in a very
short time. A recently formulated image data structure called Spiral
Architecture is the basis for speedup of processing response in our
research. The improvements of computation efficiency can be further achieved
by parallel processing on a cluster of computers.
MOTION TRACKING detects the moving objects in a dynamic scene. It is a key step in computer vision since it is the result of a successful image processing and segmentation stage, and it has a strong influence on the subsequent image analysis and recognition tasks. Its importance has generated a lot of interest among the image processing community, and has been and is the focus of numerical activities by several research groups worldwide. In an image processing application such as traffic scene analysis, it is clear that no traffic data can be extracted before the vehicle has been detected. This could be a complicated problem when real-time image analysis for real-world traffic analysis is considered.
IMAGE ANALYSIS examines image data in order to solve vision problems. It involves two subtopics: Feature Extraction and Pattern Classification. The object Edge Detection and Contour Extraction are two types of feature extraction, and Object Recognition and Object Matching are those of pettern classification.
EDGE DETECTIONextracts and localizes points (pixels) around which a large change in image brightness has occured. The performance of higher level processes such as extraction of object contours and object recognition rely heavily on the correctness and completeness of edges. Noise produced by imaging and sampling processes causes the majority of problems in edge detection. There are two classes of edge detection algorithms with noise smoothing. One of these classes is based on regularization, which is achieved by imposing smoothness constraints on the solution of edge points in various forms such as minimizing an energy functional. Another class of edge detection algorithms employ various noise smoothng processes before the actual detection procedure. Noise smoothing can be achieved by a low pass filter which is a convolution with a kernel. Multi-scale edge detection based on a Gaussian kernel has been an active research area in recent years.
GAUSSIAN THEORY for edge detection provides not only good performance of detection but also accurate localization of edges. The basis of the Gaussian theory is the multi-scale representation of the image signal. It has been proven that the Gaussian kernel is the unique kernel which processes a number of special properties on the multi-scale rpresentation.
CONTOUR EXTRACTION is frequently used in viaual information processing. The contour of an object contains relevant information about the shape of that object. Object contour extraction from given 2-D digita images is an important step for feature-based object recognition. There are two main streams of research for contour extraction of a 3-D object in a 2-D image: Object Contour Following and Multiple Step Contour Extraction.
OBJECT RECOGNITION is a process of recognising a 3-D object in a 2-D image. One of the most fundamental methods of detecting an object is matching using templates or models. This method is called model-based object recognition. In a model-based object recognition process, each model representing an known object is compared to objects in an image. There are two issues involved in object recognition. One is concerned with the representation of objects. The other one is how to detect objects using matching, given a representation scheme. Object representation provide ways of describing object properties of features. Object matching, which measures the degree of similarity between two object sets that are superposed upon one another, is a method commonly used for object recognition.
SPIRAL ARCHITECTURE is inspired from anatomical considerations of a primate's vision system. It is a relatively new image data structure. In the Spiral architecture, an image is represented as a collection of hexagonal picture elements. The distribution of cones on the primate's retina provides the basis of the Spiral architeucure. The geometrical arrangement of cones can be described in terms of a hexagonal grid. The importance of the hexagonal grid is that it possesses powerfully computational features that are pertinent to the vision process.
CLUSTER
COMPUTING
for computer vision is to use a set of computers (e.g., PCs) as a high
performance distributed system, which is able to handle real-time image
data. Real-time image data has quite a large bandwidth. As a result, it
is almost impossible, or it is extremely expensive for a single, centralized
system to handle multiple streams of real-time image data efficiently.
According to this consideration, in our research, we will develop a high
performance distributed real-time image processing system using personal
computers (PCs). The system consists of multiple PCs, several of which
have image capturing capability, connected via very high speed network.
We call this system a 'Cluster of PCs'. Over the next decade, clusters
will span the entire range of high-performance computing platforms. The
cost of commercially available PCs is getting cheaper and cheaper, and
their performance is getting higher. A cluster running a Unix-like operating
system such as Linux has little or no cost. Cluster computing can be at
least one or two orders of magnitude less expensive than a supercomputer
while it still provides a cost-effective solution and capability that was
not available on a workstation. Cost has not been the only issue in this
choice. Source code availability has been important to enable code modifications
which facilitate parallel computation on these systems.
