Cognitive Vision Group / Laboratory for Intelligent Systems
Dr. Roger S.
The mission of the Cognitve Vision Group is to develop systems capable of understanding the contents of images and the behavior of objects in video sequences. Applications for the research conducted in the group include surveillance, assisting people with special needs, highway monitoring and traffic control, steganography and steganalysis, analysis of astronomical observations, medical image analysis and automation. Furthermore, research in this laboratory is being applied to domains outside of imagery.
Members of the Laboratory for Intelligent Systems are conducting research in the development, implementation and testing of computer vision understanding and analysis systems and subsystems. Current projects include work in individual images and video sequences.
Research in the video sequence domain focuses on developing algorithms that are integrated into a framework with the goal of detecting, recognizing and understanding the activities of objects in a video sequence in real time. Motion detection has been a long-term problem in computer vision, especially in sequences acquired with non-stationary cameras and realistic changing light conditions. We are using both stationary and moving cameras to acquire data and investigating several different techniques to address the motion detection problem. Video sequences typically contain a vast amount of data, of which only a small subset is of interest. As a result of our everyday visual interactions and our analytical abilities, we acquire a general understanding of the environment and we can distinguish between consistent and inconsistent (novel) events based on prior knowledge. Any event that does not fit our definition of normalcy tends to be recognized and classified as a novel event. One goal of our group’s research is to develop a framework that will emulate novel detection in humans. Objects in the videos may include vehicles (cars, trucks, bicycles, etc.), individuals, groups of individuals and animals. Several techniques are being investigated that can be effective in aiding the tracking and recognizing of individuals. By automatically detecting, classifying (people, vehicles, animals, etc.) and tracking the objects their individual activities can be analyzed. This analysis can be used to detect and track unusual events and index the video stream allowing event-based retrievals from other video sequences, creating video mosaics, and video summarization.
With the explosive growth of the Internet and multimedia techniques there is an increasing interest in hiding data in digital media. Steganography is the science of hiding information in images that is not detectable to the human eye, whereas steganalysis is the detection of this hidden information. Members of the lab are developing mathematical techniques for both steganography and steganalysis
Automatic Object Tracking and Interpretation in Video Sequences (AOTIVS)
The AOTIVS project has two major objectives. The first objective of this project is to detect, track and classify both salient and non-salient objects in video sequences. The second objective is to model and interpret the object’s behavioral activities. The objects may include individuals, small groups of individuals, crowds and vehicles. Modeled behaviors will be used as a video indexing method. During the first phase of the project a stationary camera will be used to collect the video sequences. During the second phase networks of collaborating cameras and non-stationary cameras will be used.
As a result of our everyday visual interactions and our analytical abilities, we acquire a general understanding of the environment and we distinguish between consistent and inconsistent (novel) events based on prior knowledge. Any event that does not fit our definition of normalcy tends to be recognized and classified as a novel event. If such an event occurs often, the event is ignored as a consequence of habituation and is no longer considered novel. In this project, we extend the theory of habituation to include learning of what is and what is not a novel event and develop the foundations for a computational learning framework for novelty detection and understanding based on biologically inspired principles. Astronomical observations, surveillance systems, ultrasound and cardiac medical imaging are examples of systems that generate streams of video data. It is very tedious for a human analyst to observe and interpret all this data and find “novel events”. In order to manage and eventually utilize the information extracted from video, it is necessary to develop a framework that can extract meaningful semantic information such that it reflects the description understood by a human analyst. Such a framework will enable us to gain a better understanding of the issues and challenges associated with extracting a semantic understanding from unstructured video data.
Real Time Implementation of Video System
Video surveillance, medical image video processing, assisting people with special needs, highway monitoring and traffic control systems require real time video operation. In this project we are investigating two methods to improve the speed of our video processing algorithms. The first approach involves utilizing the Intel Integrated Performance Primitives and Math Kernel software libraries that runs on standard Intel Pentium processors. The second approach involves evaluating the Intel MXP processor.
Computational Model of B. Splendens’s Vision
While many computational models have been created to mimic the cortical visual processing in humans, few, if any, such models exist specifically for other species despite wide potential for practical application. This research draws from pre-existing models of the human visual cortex, biological research, and original experiments with B. splendens to construct a biologically based computational model that emulates the species’ ability to visually detect conspecifics, male or female, with the ultimate goal of providing a better understanding of Osteichthyes visual systems. Knowledge gained from this project will be applied to human biometrics.
Steganography and Steganalysis
Steganography is the processing of ‘hiding’ information in images. The goal of this project is to investigate published steganographic techniques and evaluate their effectiveness over a wide range of image conditions. A second goal is the development of steganalysis algorithms to detect if data is hidden in an image or not.