Agriculture Reference
In-Depth Information
model removes intensity from the original color value, thus minimizing the effects
of illumination variation within the canopy scene. Blob analysis differentiates large
segmented regions from small segmented regions that can represent image noise or
fruits on the backside of the tree that are out of reach of the robot. Individual fruit
selection within a fruit cluster and partial fruit occlusion is a significant challenge.
An algorithmic approach that combines edge detection with circle detection provides
a method of recognizing the individual fruits in the cluster so that the top fruit can be
harvested first. Figure 7.34 shows an acquired canopy image sample and the succes-
sive processed images. Both the red and blue regions are fruit pixels. Blob analysis
differentiated further the fruits into large (blue) and small (red). Further processing
was conducted on the large regions, whereas small regions were removed. The two
clustered fruits were successfully separated by circle detection, and their centroids
were determined. These centroidal positions would then become the basis for fruit
position localization, and the estimated destination for the visual servo control.
Because 2-D images lack range information, depth to target was measured in two
ways: (1) ultrasonic sensor gave a rough estimate of range to canopy and (2) using tri-
angulation method based on pseudo-stereo imaging. Once the manipulator reaches
(a)
(b)
(c)
(d)
FIGURE 7.34
Sample image processing for the automatic detection of orange fruits.
(a) Sample RGB image of orange canopy. (b) Segmentation of fruit from background. (c)
Filtering operation to remove noise. (d) Location of centroids of fruits.
