Predictive Robot Vision
Professor Benjamin and I designed and implemented a fast predictive vision system for a mobile robot, as part of a multi-year research project with two other universities: Brigham Young University and Fordham University. The vision system represents the robot's environment with a 3D gaming platform. Our system uses two Firewire cameras mounted in a pan-tilt base. I personally designed and built the pan-tilt mechanism, and wrote driver software in C to control it.The complete description of the system and its components is given online at: https//robotlab.csis.pace.edu/robotlab/wiki/Eyes which includes a short video showing the camera looking around. The development of this inexpensive vision system brings stereo color vision within the budgets of a much larger part of the research community.
The project presented many challenges and opportunities that students are not typically exposed to in Pace's computer science curriculum: programming in the C and C++ languages, fabricating and working with computer-controlled hardware, programming and using software libraries in a UNIX environment, and exploring and applying the theory and algorithms of computer vision.
C was used to write the driver for the custom pan-tilt unit, and much of the vision code was written in C++ vision libraries such as OpenCV and ERVsion. All development took place on Linux and UNIX systems and much effort was put into configuring a standardized Linux installation that would support research in the robotics lab for years to come. Though not typically a focus of computer science, the opportunity to design and build a low-cost alternative to the expensive stereo vision proudcts currently on the market was a welcome one. With a simple design and commodity parts, a stereo vision solution tailored to our needs, and the needs of other researchers was developed for a fraction of the cost of commercial solutions. During the execution of this project it was necessary to evaluate many commercial and open source computer vision libraries and determine the combination that met the requirements of the project. Systems from Intel, Evolution Robotics, Videre, VXL and others were evaluated in order to find an efficient solution that would provide the required features: image segmentation, stereo disparity, optical flow, and object recognition. The open-source, Intel OpenCV library was chosen to perform the brunt of the vision processing, and the Evolution Robotics vision package was chosen for its object recognition capabilities.
Information about the Student Author
Class of 2007, Major: Computer Science
Summary of Research Experience
Dissemination of Results
We have submitted a paper describing this project to the Conference on Intelligent Robots and Computer Vision: Algorithms, Techniques and Active Vision. This conference is sponsored by the International Society for Optical Engineering, and will be held in Boston in October, 2006. We are planning to continue the development of our vision system, to test it further, and write up the results for a suitable journal.
Dr. D. Paul Benjamin, Professor of Computer Science
Achtemichuk, Thomas, "Predictive Robot Vision" (2006). Eugene M. Lang Research Fellowships. 22.