A Machine Vision System for RealTime Object Recognition

1
A Machine Vision System for Real-Time Object
Recognition

• Xiuwen Liu and Washington Mio
• Florida State University
• May 2005

2
The Technology

• An automated, real-time machine vision system
  that can be trained to detect and recognize
  objects in images and videos.
• Key Elements
• New machine learning strategies
• Features used for object recognition
• Real-time execution

3
Example Face Recognition

• Recognize individuals from facial images.
• Subjects say, 10,000 specific individuals.
• Imaging device a video camera.
• Environment an airport or an office building.

4
Face Recognition
5
Training

• This is the most complex part of the entire
  system, yet simple from the users standpoint.
  Only training data for a specific task needs to
  be provided by the user.
• For a dynamic collection of targets, small system
  updates can be made by a non-technical user.

6
Training Data Subjects

• A few images of the subjects for the system to
  learn their main characteristics.

7
Training Data Background

• A large collection of random images to exemplify
  the expected background.

8
Performance Speed

• During training, the system can be tuned to
  optimize the performance-speed balance.
• Speed 15 fps with standard equipment higher
  rates with more computational resources.

9
Technology Advantages

• The technology will enable us to monitor, in real
  time, the information content of massive amounts
  of data generated by imaging devices in an
  automated manner.
• The system can screen video sequences for content
  and route to human operators only segments or
  frames that are likely to contain relevant
  information.

10
Scalability

• The system is designed to be able to operate with
  a large database of objects.
• Recognition workflow is structured as a cascade
  of simpler decisions. Doubling the number of
  subjects, on average, only requires the addition
  of a single layer of decisions to the cascade.

11
Adaptability

• Features utilized can discern objects based on a
  variety of characteristics.

12
Recognition Engine Tests
Recognition rates () 1,000,000 possible
features
13
State of the Art

• We are not aware of any effective solution to the
  automated, real-time object recognition problem
  currently available in the market.
• A strategy for real-time frontal face detection,
  not recognition, has been developed by Viola and
  Jones. It does not apply directly to other types
  of objects.

14
State of the Art (contd)

• Several proposed recognition strategies utilize
  somewhat unreliable features.
• Systems not adaptable to other tasks. This is
  important in applications that involve elements
  such as metal, roads, concrete, vegetation, sand,
  skin, hair, etc.

15
Potential Commercial Impact

• Automated object recognition is just starting
  to transition from research to product
  development. To our knowledge, effective
  solutions to this important problem are currently
  not available in the market.

16
Potential Commercial Impact

• Facial recognition for biometric identification
  authentication security, access control, law
  enforcement.
• Non-invasive surveillance with automated analysis
  of imagery generated by video and CCTV cameras
  and other devices homeland security.

17
Potential Impact (contd)

• Automated analysis of imagery generated by radars
  for homeland security and military applications.
• Face recognition for monitoring and locating
  personnel.
• Face recognition for teleconferencing and
  human-machine interaction.

18
Target Industries

• Government agencies identification and
  surveillance for homeland security.
• Security firms biometric identification and
  authentication, locating and monitoring
  personnel.
• Computer and communication human-machine
  interaction

19
The Scientific Team

• Xiuwen Liu
• Assistant Professor, Computer Science
• Florida State University
• Washington Mio
• Professor, Mathematics
• Florida State University

20
Xiuwen Liu

• Ph.D., The Ohio State University, 1999.
• Postdoctoral positions Ohio State University.
• Expertise computer vision, image processing,
  machine learning, artificial intelligence.

21
Washington Mio

• Ph.D., Mathematics, New York University, 1984.
• Postdoctoral positions NYU, Cornell U., and U.
  of Pennsylvania.
• Expertise computer vision, shape analysis,
  pattern recognition and topology.

22
Recent Funded Projects

• A Laboratory for Real-Time Computer Vision
  Applications, Defense University Research
  Instrumentation Program (ARO-DURIP). Liu, Mio
  collaborators, 2004-06.
• Stochastic Shape Analysis for Recognizing and
  Tracking Objects in Images and Videos, NSF and
  IC, Approaches to Combat Terrorism Program. Mio,
  Liu collaborators, 2003-04.

23
Recent Projects (contd)

• Research in Statistical Shape Theory for Image
  Understanding, Army Research Office. Mio, Liu
  collaborators, 2004-07.
• Seeking Optimal Representations and Classifiers
  for Image-Based Recognition, National Science
  Foundation. Liu collaborators, 2003-06.

24
Technology Status

• Proof of Concept
• all components tested extensively
• recognition engine capable of identifying a large
  variety of objects
• very promising results in small scale facial
  recognition experiments.

25
Technology Status (contd)

• Intellectual Property provisional patent
  application has been filed.
• Commercial Activity just starting to gauge
  commercial potential.
• Weaknesses large scale testing with live video
  and graphical user interfaces are needed to
  establish commercial viability. Addressed in next
  phase.

26
Plans for Next Phase

• Full integration of components, development of
  graphical user interfaces and a prototype.
• Estimated cost lt 100,000.
• Time needed just a few months.

Ready to leave the university?
27
Point of Contact

• Washington Mio
• Department of Mathematics
• Florida State University
• Tallahassee, FL 32306-4510
• Phone [phone]
• Email mio_at_math.fsu.edu