Visual Systems for Robots

by RNDr. Andrej Lúčny, PhD. (MicroStep-MIS, Ltd., Slovakia) presented at the Robotic Summer School 2010

• download accompanying workfiles Media:visual.zip
• unzip (recommended to C:\Work)

Part 1: Training on input image mainpulation

• start ImageJ
• open pic1.jpg
• Image / Color / RGB Split Split Channels
• apply Image / Lookup Tables / Red to pic1.jpg (red)
• apply Image / Lookup Tables / Green to pic1.jpg (green)
• apply Image / Lookup Tables / Blue to pic1.jpg (blue)
• close everything
• open pic1.jpg
• apply Image / Type / 8bit to get gray image
• apply Process / Binary / Make Binary to get binary image

• start Octave or Matlab
• change directory to directory with this training using the cd 'C:\Work\part1'
• launch tr1

Part 2: Training on 2D processing

• start ImageJ
• open pic1.jpg
• turn it grayscale: Image / Type / 8bit
• descrease noise by threshold: Image / Adjust / Threshold
  • select Black & White from menu
  • move with low and high end of range to emphasize the seen objects
  • Apply and close the threshold window
• make the image binary: Process / Binary / Make binary
• if the picture contains now black object on white background, change it to white object on black background: Image / Lookup Tables / Invert LUT
• try to eliminate holes in object by Process / Binary / Dilate and Process / Binary / Erode
• try Process / Binary / Skeletonize

• restart ImageJ
• open pic1.jpg
• start segmentation to 2 colors by Plugins / Segmentation / k-means clustering (select two colors)
• return to the opened pic1.jpg and three times apply Process /Smooth then again perform the same segmentation
• compare the two segmented images and select better

• restart ImageJ
• open pic2.jpg
• apply Process / Filters / Gaussian Blur / 1.0
• apply Process / Find edges
• apply Process / Binary / Make Binary
• apply thinning by Process / Binary / Skeletonize

• start Octave or Matlab
• change directory to directory with this training by cd 'C:\Work\part2'
• launch tr2

Part 3: Training on 3D processing

• start ImageJ
• open pic3.jpg, depthX.jpg and depthY.jpg

There is box and figure on the image pic3.jpg -- you can see that these images has the same resolution.
depthX represents X coordinate of surface point visibile at the particular pixel
depthY represents Y coordinate

How many times further is the box than the figure ?

(Hint: use ImageJ to display coordinates of the proper surface points and find distance from camera from depth images)

(Hint2: The correct answer is half of very known and famous number)

Part 4: Training on recognition

• start Octave or Matlab
• change directory to directory with this training by cd 'C:\Work\part4'
• launch tr4

This program recognizes orange color of ball and turn the image pic4.jpg to pic5.jpg

• Copy Hough_Circles.class and Hough_Circles.java into

C:\Program Files\ImageJ\plugins\ (or similar path)

• start ImageJ
• open pic5.jpg
• turn it to binary: Process / Binary / Make Binary
• Process / Binary / Dilate
• Process / Binary / Skeletonize
• Plugins / Hough Cricles / ask for 1 circle from 30 to 100 and the program should caclulate center of the circle

What will happen if we do not decrease number of white points by skeletonization ?