Robotic surveying of fruit plants
A method of machine vision includes identifying contours of fruits in a first image and a second image and performing two-way matching of contours to identify pairs of matched contours, each pair comprising a respective first contour in the first image that matches a respective second contour in the second image. For each pair of matched contours, a respective affine transformation that transforms points in the respective second contour to points in the respective first contour is identified. The second image is mapped to the first image using the affine transformations to form a composite image and the number of fruits in the composite image is counted.
1. A method comprising:
receiving a plurality of images;
forming a plurality of composite images from the plurality of images, each composite image formed by mapping pixels from multiple images onto the composite image, wherein forming the plurality of composite images comprises:
comparing a successive image to a first image of multiple images in a current composite image wherein the first image is the first image added to the current composite image;
adding the successive image to the current composite image when at least one contour in the successive image matches a contour in the first image; and
setting the successive image as a new composite image when none of the contours of the successive image matches a contour in the first image;
counting fruit in each composite image to form a fruit count for each composite image; and
summing the fruit counts to form a total fruit count.
2. The method of claim 1 wherein comparing the successive image to the first image comprises:
identifying contours in the first image;
identifying contours in the successive image;
selecting a contour in the first image;
searching for and finding a matching contour in the successive image that matches the selected contour;
searching for and finding a first image matching contour in the first image that matches the matching contour of the successive image; and
determining that the first image matching contour is the selected contour.
3. The method of claim 2 wherein identifying contours in the first image and the successive image comprises filtering pixels based on color values for the pixels to locate pixels that are likely to represent fruit and identifying contours within the located pixels.
4. The method of claim 1 wherein counting fruit in a composite image comprises applying a Gaussian mixture model to the composite image.
5. The method of claim 1 wherein forming a composite image comprises determining a separate affine transformation for each of a plurality of contours in one of the plurality of images.
6. A system comprising:
an unmanned aerial vehicle having at least one camera that captures multiple images of an agricultural plot;
a server that receives the captured multiple images and processes the captured multiple images to count fruit in the agricultural plot, wherein the server processes the captured multiple images by:
forming a plurality of composite images, each composite image formed by mapping a respective plurality of the captured multiple images to a first image of the composite image wherein forming a plurality of composite images comprises comparing a first image of a composite image of a plurality of captured images to a second image and creating a new composite image from the second image when no contour in the second image has a corresponding contour in the first image;
counting fruit in the composite images to form a separate count for each composite image; and
summing the separate counts to form a total fruit count.
7. The system of claim 6 wherein determining if a contour in the second image has a corresponding contour in the first image comprises searching for a corresponding contour in the first image and after finding a corresponding contour in the first image, searching the second image for a contour corresponding to the corresponding contour of the first image.
8. The system of claim 6 wherein counting fruit in a composite image comprises applying a Gaussian Mixture Model to the composite image.
9. The system of claim 6 wherein mapping a captured image onto the composite image comprises determining a respective affine transformation for each of a plurality of contours in the composite image, using the affine transformation to identify points of correspondence between the captured image and the composite image and mapping the captured image onto the composite image based on the points of correspondence.
10. The system of claim 9 wherein mapping the captured image onto the composite image comprises setting pixels in the captured image and the composite image to zero if the pixels are unlikely to represent fruit and not mapping pixels in the captured image to the composite image if there is a non-zero pixel in the composite image where the pixel in the captured image would be mapped to.