Location of apples in trees using stereoscopic vision

• An apple recognition algorithm based on color features is presented.
• The fruit shape features are extracted using random ring method.
• A matching algorithm based on area and epipolar geometry is presented.

In order to design a robot which can automatically recognize and locate apples for harvesting, a machine vision system was developed. Three algorithms used in the vision system to recognize and locate apples were described in this study. An apple recognition algorithm with color difference R − G and color difference ratio (R − G)/(G − B) was presented. If a pixel met R − G > 0 and (R − G)/(G − B) > 1, either the pixel was identified as an apple, else the pixel was background. The fruit shape features were extracted from contour images based on random ring method (RRM). A matching algorithm based on area and epipolar geometry was discussed to locate the apples. The apples with similar areas were matched according to the principle of ordering constraint by calculating the maximum value of cross-correlation function of vertical projections. The experiment results showed that the proposed recognition method could eliminate the influences of shade and soil. Over 89.5% of fruits were successfully recognized from 160 tested images. The circle centers and radii were extracted precisely based on the random ring method. The errors were less than 20 mm when the measuring distance was between 400 mm and 1500 mm.