Camera Calibration Using Known Object

1
Camera Calibration Using Known Object

• Computer Vision Project
• Spring 2000
• Yongjik Kim Joohi Lee

2
Goal

• Compute parameters of 2 cameras fixed on the HMD
• Use a box whose dimension is known
• Calculate
• Intrinsic parameters of each camera
• Transform between 2 cameras

3
Step 1. Two Images of a Box

• Picture (left right view)

4
Step 2. Edge Detection

• Generate edge magnitude and angle information
• (Angle information uses Gaussian filtering)

5
Step 3. Hough Transform

• Angle step 3?, Position step 2 pixels

6
Step 4. Find Edges

• Filter Hough Transform map with Gaussian filter,
  and find only local maxima

7
Step 5. Line Re-fitting

• For each line found in step 4
• find corresponding points in the band region
  along the line, and use least square fit to
  refine edges (and suppress duplicate edges)

New line
Old line
8
Before Line Re-fitting
9
After Line Re-fitting
10
Step 6. Intersection
11
Step 7. Correspondence

• For the first L-R pair, manually select which
  point in image corresponds to which vertex in
  world coordinate.
• After that
• Select 7 points
• Convex hull
• Use symmetry of box
• Find best match to Mint already found.

12
Step 8. Calibration

• Based on Trucco 6.3
• From 3x4 Projection Matrix
• Ox Oy fx fy R T

13
Experiment

• Find in the first L-R image pair (L1, R1)
• intrinsic/extrinsic matrix of each camera
• Left-to-Right transformation matrix
• To test correctness
• we map world coordinate back to the image using
  intrinsic/extrinsic matrices.
• In both images, points match within 5 pixels!
• Nice!

14
Result of L1 Image
15
Experiment

• New images (L4, R4) taken
• Compute extrinsic matrix of L4
• By using L-to-R matrix already found, compute
  extrinsic matrix of R4
• -gt we guess image coordinates of R4
• - from image data of L1, R1, and L4.

16
Result of L4 points on R4

• R4
• L4

17
Result

• Nice!