apa 6905 instrumentation for research in movement science school of human kinetics mario lamontagne...

APA 6905 INSTRUMENTATION FOR RESEARCH IN MOVEMENT SCIENCE

SCHOOL OF HUMAN KINETICSMario Lamontagne PhD

Vicon Calibration Vicon Calibration Vicon Calibration Vicon Calibration

David GrohUniversity of Nevada –

Las VegasDepartment of Kinesiology



Purpose of CalibrationPurpose of CalibrationPurpose of CalibrationPurpose of Calibration

To assist in the reconstruction 3D positions of captured markers from 2D camera images

To establish an absolute reference of origin and orientation



Earlier Calibration MethodsEarlier Calibration MethodsEarlier Calibration MethodsEarlier Calibration Methods

Direct Linear Transformation (DLT)• Abdel-Aziz YI and Karara HM (1971)

Non-Linear Transformation (NLT)• Dapena J, Harman EA, and Miller JA (1982)



Direct Linear TransformationDirect Linear TransformationDirect Linear TransformationDirect Linear Transformation

Two or more 2D views (cameras) of a control object (cube or cage) to solve for camera parameters relating the image space (external) to the object space (internal)

Both reference frame and the units of measurement are defined by the known 3-D point coordinates on the control object (Hinrichs).

Time consuming Costly



Non-Linear TransformationNon-Linear TransformationNon-Linear TransformationNon-Linear Transformation

Proposed by Dapena, et al. Also uses control object (pole) Specific point coordinates on the object are not necessary

Internal parameters calculated separately



DLT Calibration AccuracyDLT Calibration AccuracyDLT Calibration AccuracyDLT Calibration Accuracy

Chen L, et al. (1994)• Five groups of control (known) points (8,12,16,20,24)> Six configurations each

• Results: > accuracy increased as control points increased

> Best accuracy achieved with even distribution

> Accuracy decreased as distance from control region increased



NLT vs. DLT NLT vs. DLT Calibration Accuracy Calibration Accuracy

NLT vs. DLT NLT vs. DLT Calibration Accuracy Calibration Accuracy

Hinrichs RN and McLean SP (1995) Purpose:

• to compare the accuracy of the NLT vs. DLT methods

• To compare accuracy with and without extrapolation



MethodsMethodsMethodsMethods

DLT control object• Cuboid, 27 strings hanging from ceiling

• Two defined control volumes– Entire volume (185cm x 490 cm x 213 cm)

– ~25% of entire volume

• Four groups of control points (16, 24, 40, 60)



MethodsMethodsMethodsMethods

NLT control volume• Similar dimensions to large DLT volume• Single 223cm pole with six marks

> Eleven vertical positions, four horizontal positions

> Encompassed the control volume

Two cameras Mean absolute error – calculation of the resultant difference between the predicted and surveyed locations of selected points from the DLT control object (Hinrichs)



ResultsResultsResultsResults

Smallest resultant errors recorded with standard DLT re-predicting control points

Errors in re-predicting control points increased as control points increased

Errors in predicting non-control points decreased as control points increased



DynaCal3DynaCal3DynaCal3DynaCal3

Vicon Workstation’s current calibration method

Replaced DC2 in early 2001 Implements reconstruction algorithm Provides System information:

• Positions and orientations of cameras wrt each other

• Focal length of lenses• Lens distortion• Absolute reference origin of lab



Calibration ProcedureCalibration ProcedureCalibration ProcedureCalibration Procedure

Dynamic calibration • Wand dance• Determines camera and lens variables

• Processed first



Calibration ProcedureCalibration ProcedureCalibration ProcedureCalibration Procedure

Static calibration • L-frame• Determines origin of lab space (absolute reference point), only

• Depends on successful dynamic calibration

• Processed last



DefinitionsDefinitionsDefinitionsDefinitions

Residuals – numbers that indicate the quality of calibration

Linearization – the process of correcting for lens distortion

Bundle Adjustment – an algorithm that performs the linearization process by optimizing the camera parameters to give the best re-projection error



Dynamic CalibrationDynamic CalibrationDynamic CalibrationDynamic Calibration

Performed first Algorithm presumptions

• Known distance between wand markers

• No additional markers moving within the Volume during calibration

• 2D marker images seen by different cameras originate from the same 3D positions




Algorithm first looks at camera pair with the greatest overlap.

Accumulation of additional cameras initiate optimization algorithm to determine best fit.

Bundle Adjustment performs linearization process to correct for lens distortion.




Dynamic calibration algorithm uses known distance between markers to establish the scale of the volume.

This distance constant is specified in the Calibration Reference Object (CRO) file.




Wand wave (“wand dance”) tips:• Avoid breaking cameras’ views with one’s body.

• Position body so the wand can be seen by as many cameras as possible.

• Move around the entire volume to give all cameras an equal opportunity to see the wand wave.

• No need to collect more than 1000 samples



Static CalibrationStatic CalibrationStatic CalibrationStatic Calibration

Establishes absolute spatial reference (origin)

Utilizes L-frame with two axes• Tripleton – three markers• Singleton – one marker



Static CalibrationStatic CalibrationStatic CalibrationStatic Calibration

Tripleton• Establishes direction of first (X) axis

Singleton• Establishes position of origin• Combines with the first axis to establish the second axis

• Resultant cross product is the third axis

• Origin specified in CRO file



ResidualsResidualsResidualsResiduals

Individual cameras• A measure the accuracy of one camera with respect to all other cameras

• Established by reconstructing wand marker positions using all cameras except the one for which the residual is being calculated> Camera in question is non-contributory to its own 3D reconstruction residual

> New with DC3



ResidualsResidualsResidualsResiduals

Individual cameras (cont.)• The distance between the reconstructed markers image and the camera’s own image averaged across all available samples

• High camera residual = 2D contribution that tends to be less accurate than the other cameras



Residuals (cont.)Residuals (cont.)Residuals (cont.)Residuals (cont.)

Mean Residual – average of individual residuals

Residual Range – the highest and lowest residuals

Wand Visibility – the average percentage of the wand wave that contributed to each camera’s calibration

Static Reproducibility – a measure of the accuracy of the reconstructed positions of the L-frame markers as compared to the CRO file



Common Reasons for Failed Common Reasons for Failed CalibrationsCalibrations

Common Reasons for Failed Common Reasons for Failed CalibrationsCalibrations

Inappropriate wand wave• Too fast (low frequency)• Solution: slow wand wave

Poor camera positioning• Inadequate overlap• Solution:

> adjust cameras> May need to customize wand wave

Background static noise• Reposition cameras• Adjust camera sensitivities (can be turned up after calibration



ReferencesReferencesReferencesReferences

Dapena J, Harman EA, and Miller JA. (1982) Three-dimensional cinematography with control object of unknown shape. J. Biomechanics 15:11-19

Chen L, Armstrong CW, and Raftopoulos DD. (1994) An investigation on the accuracy of three-dimensional space reconstruction using the direct linear transformation technique. J. Biomechanics 27:493-500



ReferencesReferencesReferencesReferences

Hinrichs RN, McLean SP. (1995) NLT and extrapolated DLT: 3-D cinematography alternatives for enlarging the volume of calibration. J. Biomechanics 28:1219-1223

apa 6905 instrumentation for research in movement science school of human kinetics mario lamontagne...

Documents

orientation slide

extrapolation slide

control object hinrichs

control object cube

control region

groups of control points

dlt methods

best accuracy