Finding the Tracking Features in the Video Stream Requires Image Processing

Two Genres of Processing were Developed for This Research

  • Primitive Image Processing
  • Statistical Image Processing

 


Primitive Image Processing

 

Primitive Image Processing Functions

  • Basic Image Manipulation
  • Minor Image Fixes
  • Quick Running
  • Non-Robust
 
Visual Results of Primitive Processing Functions

 


Statistical Image Processing

 

Statistical Image Processing

  • More Powerful
  • Account for a Wide Range of Circumstances
  • Slower-Running
  • Typically an All-In-One Solution

Statistical Algorithms are Based on the RGB Color Cube Concept

  • Digital Color Can be Represented as a 3d Cube in Space
  • Origin is Black (i.e. R=G=B=0)
  • For 24-bit Color, Red, Green, and Blue Axes run from 0 to 255
  • Any of the 16 Million Colors Can be Defined as a Spatial Coordinate in Color Cube

Statistical Algorithms Used:

  • Color Range
  • Color Direction
  • 3D Gaussian Color Distribution
  • 2D Normalized Gaussian Color Distribution

 


Color Range Classifier

 

  • Simplest
  • Fastest Running
  • Used for Simple Color Distributions
The color range classifier determines a box in RGB space.

   
Color Range Distribution Shape in RGB Space

The color range classifier was tested on determining a yellow lid from its surroundings.

The image on the right shows the original image.

 
   
Original Image of Yellow Lid

The image on the right shows the same yellow lid image with the classifier results displayed.

Red areas were added by the software to indicate positive classifier hits.

 
   
Color Range Classifier Results

This is the data distribution in RGB color space.

All the points in the graph indicate locations of pixels in the image. The white points indicate correctly classified pixels. The black points indicate correctly unclassified pixels. The blue points are yellow lid pixels missed by the classifier.

 
   
Yellow Lid's Color Distribution in RGB Space

 


Color Direction Classifier

 

  • Searches for Vector Information in RGB Color Space
  • Designed to Find Colors Even in the Presence of Shadowing
  • Used When a Color Is Unique in a Scene, but Lighting Conditions are not Consistent

Color Direction's distribution in RGB space is a conic section.


   
Color Direction Distribution in RGB Space

The color direction classifier was tested on a simulated warehouse environment. The goal was to detect the biohazard symbols on the barrels.

The original image is shown to the right.

 
   
Original Image of Biohazard Barrels

This image shows the color direction classifier results overlaid on the orignal image.

Classifier results are shown in yellow.

 
   
Color Direction Classifier Results
This graph shows the distribution of the color in the image. The white points indicate correctly classified pixels in the image. The black points indicate the correctly unclassified pixels. The red points indicate pixels that were inaccurately classified. The blue points indicate pixels that were missed by the classifier.  
   
Biohazard Barrel's Distribution in RGB Space

 


3D Gaussian Classifier

 

  • Most Common Image Classification Method
  • Based on Data Variance and Acceptable Error

Classifier Distribution is an Ellipsoid in RGB Space


   
3D Gaussian Distribution in RGB Space

The 3d Gaussian Classifier was used on a local road to attempt to distinguish the road lines from ambient data.

The original image is shown to the right.

 
   
Original Image of Street

This image shows the 3d Gaussian classifier results overlaid on the orignal image.

Classifier results are shown in yellow.

 
   
Color Direction Classifier Results
This graph shows the distribution of the color in the image. The white points indicate correctly classified pixels in the image. The black points indicate the correctly unclassified pixels. The red points indicate pixels that were inaccurately classified. The blue points indicate pixels that were missed by the classifier.  
   
Street's Distribution in RGB Space

 


2D Gaussian Classifier

 

  • Hybrid of 3D Gaussian and Color Direction Methods
  • Requires that Color Information be Multiplied by Co-Planar Color Normalization Vectors
  • Removes Intensity Information from Color
  • Good at Compensating for Non-Consistent Lighting Conditions

Classifier Distribution is an ellipse in 2D normalized color space. The 2D color space falls into a planar triangle with each primary color at a tip.

Color triangle is shown to the right.


   
2D Normalized Color Space

The 2d Gaussian Classifier was used on the same yellow lid from the color range classifier.

The original image is shown to the right.

 
   
Original Picture of Yellow Lid

This image shows the 2d Gaussian classifier results overlaid on the orignal image.

Classifier results are shown in red. Notice that the results of this classifier are much more accurate than the color range classifier.

 
   
2D Gaussian Classifier Results
This graph shows the distribution of the color in the image. The white points indicate correctly classified pixels in the image. The black points indicate the correctly unclassified pixels. The red points indicate pixels that were inaccurately classified. The blue points indicate pixels that were missed by the classifier.  
   
Street's Distribution in RGB Space