Matt's Research
My main research interests are machine vision and learning, especially applied to robotics. You will find brief summaries of my current projects below, and a complete 4-page research statement is available here (in PDF).
Structure Learning for Autonomous Mobile Robots
Intelligent systems and robotics technology together stand poised to revolutionize the way human beings live and work. In recent years, robots have begun to leave the laboratory; in the next decade, we will see explosive growth in industrial and consumer-oriented robotics products. Going far beyond toys for our entertainment, robotics technology will rapidly expand to keep us safe, keep us healthy, and eliminate dangerous and tedious tasks from our lives.
The main factor driving this growth is the semiconductor industry's relentless expansion of cheap compute power. We are already seeing relatively sophisticated machine intelligence applications being delivered on hardware costing less than 10 USD per unit in volume. The amount of compute power available for any fixed price will continue to grow exponentially for the foreseeable future.
However, although massive amounts of cheap compute power will be available to robotics applications in the near future, the algorithms we have for exploiting this massive compute power are sorely lacking. When compared to humans (or even rats), modern technology is most glaringly deficient in perception of the surrounding environment. The gap in visual perception abilities between animals and machines is perhaps the greatest obstacle robotics practitioners face today.
Machine vision research aims to close this crucial gap. In particular, for most mobile robot applications, the crucial unsolved machine vision problem is structure learning, wherein the system must determine the 3D structure of the nearby environment from one or more moving cameras. Unfortunately, after more than 20 years of research, we only have rudimentary, poorly-performing structure learning algorithms.
Over the last 15 years or so, many areas of artificial intelligence have been wholly transformed by the introduction of statistical learning theory. For machine vision, this transformation has only barely begun, perhaps mainly due to the high computational demands of statistical inference algorithms when operating on extremely high-dimensional data (images). But with the advent of inexpensive 1000 MFLOPS machines, the time is ripe for research at the intersection of machine learning and machine vision.
The general framework of my research is to 1) formulate 3D structure learning problems as problems of statistical inference, 2) specify statistical models appropriate for the problem at hand, and 3) devise efficient algorithms for inference under said statistical models. One example in my current work is Bayesian estimation of feature correspondences in trinocular stereo images. In stereo, the goal is to find corresponding features in two or more cameras (three cameras in the case of trinocular stereo) with known calibration parameters, then use triangulation to determine the distance to the identified feature. Once a robot knows which points correspond to each other in a set of images, it can construct a 3D representation of the nearby environment. But finding those correspondences is a difficult problem, and current existing approaches are neither accurate enough, robust enough, nor efficient enough for real-time use by autonomous mobile robots. I believe that the statistical learning approach will lead to new, effective solutions to this and other difficult problems in robot visual perception.
Visual Tracking
Beyond understanding the static 3D structure of the world, many future applications in robotics will require the ability to keep track of the dynamic location of various of objects through time. I am interested in methods for real-time 2D and 3D visual tracking, especially of humans and their body parts. I am currently working on a human hand tracking system with Nyan Bo Bo at SIIT. We envision a system that will be useful not only for robots interacting with human beings, but also for surveillance applications such as shoplifting or pickpocketing detection, which require observing human hand gestures.
Network and Systems Security
I am also interested in research at the intersection of systems security and artificial intelligence. As we come to rely more and more on the Internet for our commercial, governmental, and recreational activity, we become increasingly susceptible to identity theft and attacks on our privacy. As the techniques for attacking our systems become increasingly sophisticated, we need new methods to prevent and react to those attacks. I am interested in new approaches to security from the perspective of intelligent systems. I am currently working with Chanathip Namprempre at the Faculty of Engineering, Thammasat University, on the use of CAPTCHAs (Completely Automated Public Turing tests to tell Computers and Humans Apart) to improve network security. For more details on this project, see our CAPTCHA page.
