Ben Schneider

Zack Sorenson

Neil Agarwal

Bailey Zhang

Vision Recognition Outline

What is Vision Recognition?

           Definition

           Computational Vision

           Attempts to understand visual perception and understanding Machine Vision

           Attempts to create practical application for computer vision

How does it relate to AI?

           Creating useful autonomous agents

                     -Robocup is the tournament where they have dogs play soccer - obviously
                     vision is useful here.  They have a goal to by 2050 create bipedal robots
                     capable of beating the best human soccer players.

                     -Asimo is a bipedal robot which has cameras that allow it to interact with
                     its surroundings.

           Human/computer interfacing

           -Gesture and gait recognition.

                     -GRACE is a robot that uses vision to recognize her surroundings.
                     GRACE attended a conference, navigating her way through a lobby and up
                     through and elevator.  GRACE also has a newly installed binocular vision
                     system that allows her to recognize human gestures.

                     -Recognition of handwritten text and drawings.

           Passive observation & analysis

           -For example observing a landscape and looking for changes over time.
-Traffic and surveillance cameras that automatically identify passing objects.

           Understanding biological vision

                     -"Spatial reasoning" is something we all talk about but no one really
                     understands.

What's being done in Vision Recognition?

           Penn State Computer Vision Lab

                     Hand gesture recognition as supplement to speech recognition.

                     Text recognition and clarification in video

Current Implementation

           Optical Character Recognition (OCR)

                     Types of OCR

                                   Character-based word recognition

                                   Segmentation-based word recognition

                                   Word-shape recognition

                     Current Limitations of OCR

                                   Underlines in text

                                   `I' and `i' confusions

                                   Distortion of `e'

                     Examples of OCR software

                                   Calera Wordscan Plus 1.0

                                   Caere Omnipage Professional 3.0

                                   Xerox Imaging Systems AccuText 3.0

           Scarecrow

                     - Floating, autonomous robot platform powered by solar panels

                     - Searches for birds using its vision system

                     - Goes around, when its senses birds, bumps them or shoots water at them

                     - Floats in catfish ponds to protect the stock which is frequently depleted
                     by hungry birds

Successes

           Machine Vision is being developed to control the quality of products, detecting
           defects automatically

           The technology is reaching a point where people can be tracked in 3D, leading to
           more robust security systems.

           And for your entertainment, Machine vision is being developed to track actual
           movements, for more realistic and exciting games

           Plus, more specific examples

Problems and setbacks

           Finding 3D objects in 2D images

                     -Researchers are trying out binocular vision systems, but they work only
                     for hardware, not software.

           Making raster images usable for a computer

                     -Usually all the computer has is a bunch of pixels that it has to convert into
                     meaningful shapes.

                     -The general steps for recognizing an object within an image are:

                     -Distinguish the object from its surroundings

                     -Find the shape of the object

                     -Recognize the object as a chair, apple, etc.

                     -Mathematical algorithms like edge detection and averaging are used

           Incomplete understanding of human vision

                     -Even as AI researchers look into computer vision, biologists continue to
                     look into human and animal vision

                     -AI people genereally like to look at vision as a bunch of discrete, isolate
                     processes that each do one thing, but in biological vision systems the
                     different systems may overlap and meld together

Future of Vision Recognition

           PC's Recognizing Users (Microsoft Project)

                     - When a user sits in front of the computer, it will bring up his work, etc.

                     - PC equipped with a camera on top

                     - Long term: aiming for a system that can recognize eye movements, so
                     that software could be controlled by eye movement

                     - Only a couple years from reaching mass market?

           Augmented Reality

                     Used for surveying

                     More exciting gaming

                     Tourist "goggles"