Segmentation of Apple Fruit from Video via Background Modeling
The cost of labor for orchard tasks currently accounts for over 50% of orchardists' costs.  In addition, the seasonal nature of many orchard tasks and high volume of workers needed results in a shortage of labor, especially for harvest work.  Consequently, a computer vision system to segment apple fruit in video was developed for employment in such orchard automation tasks as assessing fruit load and harvesting.  The method developed in this project, GMOG, or Global Mixture of Gaussians, uses the principles of background modeling.  The background model is learned from images of trees without fruit.  In preliminary tests, the method worked well on red and yellow fruit.
For technical details of our method, we refer to the publications at the end of this page.
The following images show the sample results of our method and videos can also be found below.  The image on the left is acquired by a digital camera mounted on an over-the-row harvester frame.  The image in the right is the result of using GMOG. 
Seq. 1: 'Dixie Red'
Seq. 2: 'Dixie Red'
Seq. 3: 'Ace Spur'
Seq. 4: 'Gold Blush'
The following table describes the image sequences.
Quantitative results from 10 ground truth images per set are listed below.
The following video sequences show the images acquired from the over-the-row harvester on the left-hand side and the right-hand side is the result from our method.
    NOTE: The .avi file is the best quality, but if this version does not play, try the .wmv version.
Sequence 1, 'Dixie Red'
Sequence 2, 'Dixie Red'
Sequence 3, 'Ace Spur'
Sequence 4, 'Gold Blush'
    •    Amy Tabb (work sponsored by the Appalachian Fruit Research Station, ARS/USDA)
    •    Johnny Park
A. Tabb, D. Peterson, and J. Park.  "Segmentation of Apple Fruit from Video via Background Modeling." ASABE Annual International Meeting, 2006. [pdf]