A CAMERA CALIBRATION TECHNIQUE USING TARGETS OF CIRCULAR FEATURES

1
A CAMERA CALIBRATION TECHNIQUE USING TARGETS OF
CIRCULAR FEATURES

• Ginés García Mateos
• Dept. de Informática y Sistemas
• Universidad de Murcia - España

2
INTRODUCTION

• Camera calibration estimation of the unknown
  values in a camera model.
• Intrinsic parameters.
• Extrinsic parameters.
• Calibration target object of known geometry,
  easy to detect and locate, used in calibration.

3
INTRODUCTION

• The whole procedure of camera calibration
  Heikkilä et al. 97
• Determinate a camera model.
• Control point location in the images.
• Camera model fitting.
• Image correction for distortion.
• Estimate the errors of the previous stages.

4
INTRODUCTION

• Much research has been devoted to model fitting.
• Control point location
• Design physical target structure.
• Design an algorithm for target detection and
  location.
• Goals accuracy, robustness, efficiency,
  simplicity.

5
TARGET DESIGN

• Previous work square features.

• Typical methods use
• Edge, segment, corner detection.
• Line intersections.
• Contour following.

6
TARGET DESIGN

• Previous work dot features.

• Point features (less than 5 pixels radius).
• Centroid calculation.
• Used in photogrametry.

7
TARGET DESIGN

• Circular features. Key ideas
• Circles (ellipses) are mapped to ellipses (using
  perspective projection).
• Ellipses are the most simple shape to describe,
  detect and locate.

8
TARGET DESIGN

• Previous work based on centroid.
• Problem of perspective bias ellipse centroid is
  not necessarily the projected centroid of the
  circle.

9
TARGET DETECTION/LOCATION

• Process for detection and location of the target.
  Main steps
• Detection and location of ellipses.
• Extraction of invariant points.
• Matching with known points of the target.
• Then model fitting (DLT) is applied.

10
TARGET DETECTION/LOCATION

• Ellipse detection and location
• Image binarization.
• Threshold median value of partial histogram.
• Connected component grouping.
• Gaussian component description.
• For each region ?, ? and number of points.

11
ELLIPSE DETECTION/LOCATION
Acquired image
Binarization
Connected compo-nent grouping
Gaussian description
12
ELLIPTICAL SHAPE TEST

• Gaussian parameters ?, ?.

• Ellipse mayor and minor radius a, b
• Ellipse area SR?ab

• Radius from gaussian parameters

13
TARGET DETECTION/LOCATION

• Ellipse location is insufficient invariant
  points should be extracted.
• Feature points in a target of circles.
• Ellipse centroid is not an invariant feature
  point.
• Invariant feature points can be obtained using
  relations between coplanar circles.

14
TARGET DETECTION/LOCATION
Perspective projection

• Tangent invariance supposing perspective
  projection common tangent property remains
  invariant.

15
TARGET DETECTION/LOCATION

• Some conclusions dont held when radial
  distortion is considered.
• Dealing with distortion
• Iterative method parameter calculation/image
  correction.
• Independent estimation (and correction) of
  distortion.

16
EXPERIMENTAL RESULTS

• Tests are centered on the target
  detection/location procedure.
• Accuracy feature point location.
• Robustness defocusing and noise.
• Efficiency computation time.
• Acquisition low-cost videoconference camera
  QuickCam Pro (Logitech).
• Computer off-the-self PC, with K6 at 350Mhz.

17
EXPERIMENTAL RESULTS

• Target used in the experiments.

320x240 pixels 256 gray levels

• Manual measure to determine ground-truth
  positions.

18
EXPERIMENTAL RESULTS

• Location error vs. ellipse size in images

19
EXPERIMENTAL RESULTS

• Manual measure is insufficient.
• Accuracy of the method (using ideal images) 0.05
  pixels mean, 0.03 pixels standard deviation.

• The target was detected in 97 of the images.

20
EXPERIMENTAL RESULTS

• Robustness to defocusing and noise.

Location error vs gaussian smoothing
Location error vs. random noise
21
EXPERIMENTAL RESULTS

• Efficiency
• The main process is a connected component
  labeling algorithm.
• This requires a single scanning of the image,
  with a constant cost per pixel.
• The whole process can be made at approx. 10 Hz.

22
CONCLUSIONS

• A technique for camera calibration is proposed
  based in the use of circles as target features.
• This contribution is centered in target
  detection/location.
• Process of detection and location
• Gaussian description of connected component.
• Feature point calculation and matching.

23
CONCLUSIONS

• The method is simple and low-level, which implies
  efficiency and robustness.
• Subpixel accuracy is clearly reached.
• High robustness to noise and defocusing.
• The technique is suited for automated systems.

24
LAST

• This work has been supported by CICYT project
  TIC98-0559.
• Línea PARP web page
• http//www.dis.um.es/parp
• Muito obrigado