Camera Coordinate Calibration Overview - gabrielburnworth/plant-detection GitHub Wiki

###Let's say we have a photo with three plants in it whose locations and sizes we need to detect: 1
These example images were created with rotation to simulate a rotated camera. Background grid shown for reference.

Assumptions:

  • We always know the coordinate location of the camera, since it is CNC-controlled.
  • The image center is directly below the camera.
  • The camera is the same distance away from the calibration objects and the objects to detect.
  • The pixels in the image the camera takes are square.

Step 1: Take calibration image of known objects

With the (imaginary) camera at the same height from the subjects, we take a photo of two calibration dots of known spacing:

2
These example images were created with rotation to simulate a rotated camera. Background grid shown for reference.

We know that these objects are 1000 units apart and that they are square with our desired coordinate system.

Note: Camera coordinates are not needed for calibration. Calibration does require the orientation of the coordination objects to be known: that they are along the x-axis and that the origin is in the lower left.

Step 2: Run calibration

3

Using our known values, we determine that the camera is rotated 5 degrees from the coordinate system, and determine that the pixel to coordinate conversion scale is 1.72.

Step 3: Determine locations and sizes of "plants" in original image:

4

Using the calibration parameters determined in step 2, and the known location of the camera in this image (800, 500), we find the locations and sizes of the "plants":

X (calculated) Y (calculated) dia. (calculated) X (actual) Y (actual) dia. (actual)
300 299 102 300 300 100
898 601 104 900 600 100
99 799 104 100 800 100