lidermitinnai Jnm'nal nf

Bin.Medical
ELSEVIER International Journal of Bio-Medical Computing 41 (1996) 137-151

Optimization and smoothing techniques in movement analysis

Angelo Cappello a,*, Pier Francesco La Palombara b, Alberto Leardini b
~Department of Electronics, Computer Science and Systems, University of Bologna, Bologna, Italy
bBiomechanics Laboratory, Istituti Ortopedici Rizzoli, Bologna, Italy
Received 12 September 1994; revised 13 March 1996; accepted 20 March 1996

Abstract

This paper deals with accurate estimation of the degrees of freedom (DOF) of a body segment, starting from the
trajectories of clusters of markers acquired with a suitable measurement system. The most commonly employed
estimation procedures involve two sequential steps, a trajectory smoothing algorithm and a DOF reconstruction
routine. Three optimized smoothing and reconstruction schemes are described, analyzed and tested and their
performances are compared with each other and with those of a more traditional technique which contains no
optimization criteria. All three schemes include an iterative, weighted-least-squares DOF reconstruction routine and
a self-tuning, zero-phase-shift, 4th-order Butterworth filter. Both routines are extensively described and validated on
the basis of numerically-simulated marker trajectories. Test results, analyzed on a statistical basis, show that the use
of an optimization routine provides a visible improvement in D O F reconstruction. This performance has also been
confirmed using stereophotogrammetric data collected on a subject wearing an external fracture fixation device which
provides reference values for the bone DOF. Angular DOF estimated applying the optimized method to skin technical
clusters are much closer to the reference values than the non-optimized values. Smoothing of data further improves
the reconstruction accuracy while a far less crucial role is played by the order in which smoothing and reconstruction
routines are applied.

Keywords: Biomechanics; Human motion analysis; Degrees of freedom; Position and orientation; Optimization;
Smoothing

I. Introduction b i o m e c h a n i c a l a n d clinical studies. T h e descrip-

tion o f b o d y segment a n d j o i n t k i n e m a t i c s a n d
Q u a n t i t a t i v e analysis o f h u m a n m o v e m e n t has kinetics requires the e s t i m a t i o n o f t h r e e - d i m e n -
been recognized to be o f great significance in b o t h sional (3-D) p o s i t i o n a n d o r i e n t a t i o n o f bones
d u r i n g m o v e m e n t . T o this p u r p o s e , l a n d m a r k tra-
jectories are n o r m a l l y used, m o s t l y o b t a i n e d b y
* Corresponding address, Dipartimento di Elettronica, In- m e a n s o f s t e r e o p h o t o g r a m m e t r i c systems. In or-
formatica e Sistemistica, Universit/t di Bologna, Viale Risorgi-
mento, 2, 40136 Bologna, Italy. Tel.: + 39 51 644 3008; Fax: d e r to allow within- a n d b e t w e e n - s u b j e c t c o m p a r i -
+ 39 51 644 3073; E-mail: [email protected]. sons, the k i n e m a t i c d e s c r i p t i o n o f m o v e m e n t

0020-7101/95/$15.00 © 1996 Elsevier Science Ireland Ltd. All rights reserved

PH S0020-7101(96)01167-6
138 A. Cappello et al. / International Journal ~71"Biomedical Computing 41 (1996) 137 151

should use reference frames associated with bone correlated with and has the same frequency con-
anatomy. According to this consideration, many tent as the bone DOF. Consequently, compensa-
authors have presented experimental methods, tion for such an artefact is not possible without
aimed at investigation into lower limb kinematics, taking into account explicitly the soft tissue me-
which are based on direct placement of a certain chanics. The influence of this artefact can be
number of markers on anatomical landmarks and reduced by locating the areas on each segment
on fixation of additional markers that permit the where skin sliding, muscular contraction and soft
definition of anatomical planes or axes [1-3]. The tissue inertial effects are globally minimized, and
use of these techniques, in which the trajectories using optimization techniques for DOF recon-
of anatomical landmarks are assumed coincident struction. Preliminary results were presented in
with the ones of their respective markers, intro- this context by the authors and other researchers
duces an unavoidable constraint that prevents any [17 19].
investigations on optimal marker positioning ac- As far as anatomical frame identification is
cording to error minimization criteria. Moreover, concerned, even though clear definitions and re-
markers are usually positioned close to the joints, peatable examining procedures have been pub-
that is on areas which have proved to be the most lished [20,21] in order to minimize the uncertainty
strongly affected by skin sliding effects [4]. In this with which bone prominences are identified by
respect, the distinction between technical and external palpation, the amplitude and direction of
anatomical frames and the use of a standardized these errors depend strongly on the examiner's
experimental protocol, the 'Calibrated Anatomi-
skills and expertise.
cal System Technique' (CAST) introduced by
This work concerns problems in 3-D recon-
Cappozzo et al. in 1994 [5], permits to focus
struction of position and orientation of rigid bod-
attention uniquely on technical cluster optimal
ies during movement, starting from noisy
data collection and artefacts reduction problems.
trajectories of rigid and deformable clusters of
Precision and accuracy in estimating the six
markers. Its aim is to suggest methods to reduce
degrees of freedom (DOF) of a body segment are
the effects of the above-named error sources. As
critical features of every study concerning human
mentioned above, a large number of good results
movement. Three main error sources must be
taken into account to predict the accuracy: (i) have been achieved and published, relative to
instrumental inaccuracies, (ii) non-rigid interface instrumental error minimization by means of
between technical cluster and underlying bone, smoothing techniques on single marker noisy tra-
and (iii) difficulties in identifying the anatomical jectories, and, more recently, by using marker
frames in a repeatable way. redundancy and algorithms based on least-squares
The instrumental error, usually quantified as a methods. Smoothing algorithms are generally
fraction of the measurement volume, has a broad used to improve the signal-to-noise ratio of dis-
frequency spectrum strongly affecting velocity and placement data before differentiation. In some
acceleration estimates. Its influence on the esti- cases, smoothing and differentiation are per-
mated DOF can be minimized by using suitable formed jointly in a state space signal representa-
smoothing routines [6-8] and optimization tech- tion [6].
niques [9-13]. If rigid fixtures are used to fasten Several methods using this approach and based
the markers to the body segment, the knowledge on time-domain [6,7] or frequency-domain [8]
of the cluster's rigid geometry in addition to the techniques were proposed in the past to be em-
use of a redundant number of markers can be ployed on biomechanical data. The best results
exploited to reduce the influence of this first kind have been generally obtained when algorithm de-
of error. sign variables are optimally tuned to signal and
A far more important source of error is the noise characteristics. Comparative assessment of
relative displacement between skin markers and different smoothing and differentiation algorithms
the underlying bone [4,14-16], which is strongly have been presented in Refs. [22,23].
 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137 151 139

In order to estimate body segment position and and a local model of the cluster can be thus easily
orientation, most of the above-mentioned opti- determined. When a skin technical cluster is used,
mization techniques determine the rotation matrix the local (rigid) model can be identified as the
and the translation vector of the local to global configuration during the calibration phase or an
(laboratory) reference transformation. Unfortu- average configuration during motion. Body seg-
nately, though characterized by high computa- ment position and orientation can be estimated by
tional efficiency, rotation matrix assessment calculating the translation and rotation which,
presents at least two major drawbacks: the lack of frame by frame, match, in the best possible way,
direct estimation of the three angular coordinates the cluster model to the actual cluster as it has
and their respective accuracy and the difficulty of been observed by the measurement system. The
dealing with anisotropic measurement noise. above-mentioned position and orientation can be
The orientation vector, together with the posi- denoted by a couple of 3-D vectors that we will
tion vector of a point of the body segment, seems call T and O, respectively. The latter variable is a
to be the best numeric description of the D O F of 3-D vector, since three independent quantities are
a rigid body and provides a compact representa- sufficient to identify the orientation of a rigid
tion of the relevant information [24,25]. The nine body uniquely.
direction cosines of the rotation matrix can be A large number of symbolic representations of
easily computed from the orientation vector com- orientation are widely known [24], the most popu-
ponents. lar of which are perhaps Cardanic-Eulerian angles
The present work begins with the description of and the rotation matrix, that actually includes
an iterative algorithm which estimates the transla-
nine different elements linked by six independent
tion and orientation vectors of a moving rigid
analytical relationships.
body from noisy marker measurements through a
As the observed cluster does not normally
weighted least-squares (WLS) procedure. In D O F
match the cluster model because of noise intro-
reconstruction, both rigidity constraints and an-
duced by the measurement system and, eventu-
isotropic properties of measurement noise are
ally, the skin artefact, an estimation procedure is
taken into account to improve estimation quality.
essential to give reliable assessments of the seg-
Algorithm performance sensitivity to the number
ment DOF. The most commonly used D O F re-
of markers and to algorithm tuning is then dis-
cussed. After that, a smoothing algorithm based construction procedures use least-squares
on a fourth-order bi-directional Butterworth filter techniques [9-13].
with cut-off frequency optimization is introduced. For the purposes of this work, an iterative
Finally, the performances of four different strate- WLS procedure [26,27] was used to estimate the
gies combining smoothing and optimization tech- (6 × 1) D O F vector, Dj = [O; T], by minimizing
niques in different ways are evaluated. the weighted sum of the squared errors (along
each laboratory axis) between the measured labo-
ratory marker coordinates, ~i(i = 1.... n), and
2. Methods those predicted by the rigid body transformation,
~i(Dl) = R(O)xi + T, where R(O) is the rotation
2.1. Body segment motion reconstruction: use o f matrix corresponding to the orientation vector, O,
optimization techniques and x i (i = 1..... n) is the local coordinate vector
of the ith marker.
The motion of a body segment is usually recon- Each of the error components along the three
structed by using the laboratory trajectories of a laboratory axes can be weighted separately and,
variable number of markers (at least three of in principle, also cross-correlation between noise
which must be non-collinear) arranged in a suit- components can be taken into account. This is
able cluster. If markers are mounted on a rigid advantageous when sufficient information about
framework, their relative distances keep constant error statistics is available.
140 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137-151

Furthermore, the weighting factor for each numerical values that are inversely proportional
marker should be related to the accuracy in its to the noise variance along the corresponding
determination while cross-correlation between lo- direction. The diagonal elements of W~ are usually
cal displacements of different markers could be defined by assigning the unitary value to one of
considered to account for skin sliding errors. them and adjusting the others according to the
In order to obtain uncorrelated position and above-mentioned ratios. When either the three
orientation estimates, and without loss of general- reference directions are equally noisy or no hy-
ity, the origin of the local reference system is potheses can be drawn on disturbance statistical
assumed coincident with the cluster centroid, nature, W is the (3n x 3n) identity matrix.
Due to the non-linear relationship between the
Xi/n. model output, ~', and the parameter vector, Dr,
i=1 the WLS algorithm must be iterative. Based on
If noise introduced by the measurement system model output linearization [28], it starts from a
is additive, we can assume that for each marker reasonable initial guess of D t as, for instance, the
optimal value attributed to Df at the previous
Yi = ~i(Df) + e, (1) frame. At each iteration, the algorithm calculates
where the error mean is E{ei} = 0 and the the (3n × 6) sensitivity matrix, S OY/ODf, as-
=

covariance matrix, E{eie~} = M, is supposed to sessed according to the current parameter esti-
be known and unvarying for the different mark- mate, and updates the parameter vector by adding
ers. If we define the (3n x 1) column vectors the maximum likelihood adjustment term
= [.v~;.-..;.V,], Y = [.Vl;....;.v,] and E = [el;....;e,] ~O~. = (S~¢¢S) - ~SWC(Y - ? ) (3)
we can write ~" = ~'(Df) + E and the cost
As ~i (Dj) = R (O)xi + T and the analytical
function to be minimized can be expressed as
relationship between R and O is known, the sensi-
follows:
tivity matrix can be evaluated analytically. Vector
F(Df) = [of -- ~'(Df)]rW[Y - Y(Df)] (2) Dj progressive adjustment causes a gradual de-
crease of the cost function, F(Dy).
where W is a (3n x 3n) positive definite weighting
The algorithm, the flow diagram of which is
matrix made up by n identical (3 x 3) sub-
illustrated in Fig. 1, stops when function F de-
matrices W~ lined along its diagonal and by zero
creases below a pre-defined threshold or when its
elements elsewhere if markers displacements are
relative reduction after two consecutive steps is
assumed to be uncorrelated. If noise is normally
smaller than a predetermined threshold.
distributed and we set W s = M - ~ , the WLS
The sensitivity matrix assessment plays a domi-
estimator, I)F, coincides with the Gauss-Markov
nant role in the above described procedure as it
estimator and the procedure provides unbiased
allows to calculate the DOF estimate covariance
DOF estimates. The use of a weighting matrix is
matrix [28],
therefore helpful when it is possible to formulate
hypotheses on noise statistical distribution. COV(I)j.) = ( S T W S ) - lG2oise (4)
When noise components along the three axes of which can be regarded as a measure of the estima-
the laboratory reference system are not cross-cor- tion process accuracy. The best results are
related, M is a diagonal matrix whose elements achieved when fixing W,. = M i because in this
represent the variances along the three directions. case the WLS estimator is maximally efficient. In
More, the ratios between the variances of noise the formula above, O'2oise is the noise variance
components are often predictable. In stereopho- along the direction corresponding to the unitary
togrammetric marker position reconstruction, for value in Ws. When noise variance cannot be theo-
example, the depth coordinate, z, is noisier than retically assessed, due to unavailability of suffi-
the two other ones, x and y. In these cases, the cient information about the measurement system,
maximum likelihood criterion suggests to assign its value can be replaced by the following esti-
the diagonal elements of matrix W, = M 1 mate:
 A. Cappello et al. /International Journal q[ Biomedical Computing 41 (1996) 137-151 141

~2°i~ - n nfr i~ (5) lent rotation matrix R(O) is expressed by the

following formula [24]:
where nfr is the number of frames taken into sin8 1 - cos8
R(O) = Icos8 + ~ - A ( O ) + 82 O0 ~
account and F °p' is the optimal value of the cost
function at frame i. (6)
It is probably worth highlighting that virtually where 8 = ~ and
any parameter sets capable of representing the
orientation of a rigid body are applicable, pro- 0 -8~ 8y 1
vided that their analytical relationships with the A (0) = 8~ 0 - &.
rotation matrix, R, are known. For our aims we
have preferred the orientation (or helical) vector
L-o ox o J
t9 = [O, 8y 8z]T = 8n, where 8 is the finite is the skew-symmetric matrix with axial vector O.
rotation angle about the axis defined by the unit The parametrization we employed lacks some
vector n, necessary to make two reference systems common vectorial properties such as additivity,
coincide. The attitude vector, 19, is normally ex- commutativity and associativity and, in addition,
pressed through its components relative to a its components rarely have an evident physiologi-
cartesian reference system. The unique, analytical, cal and clinical interpretation. Despite that, it
non-linear relationship between O and the equiva- often represents a convenient choice, as it offers
some major advantages when compared with
other representations (e.g., Cardanic-Eulerian an-
/ Initialization gles). These mostly concern its being always well-
Dr=Dfo defined and quite reliable, due to the moderate
cross-correlation existing among its components
"4, [24,29]. Further, when the Gauss-Markov esti-
mate O is assessed and both errors and marker
Model output
^ ^
Y=Y(D:) distribution can be assumed isotropic, a useful
Dr-- Df + ~Df Sensitivity matrix formula that allows to estimate the orientation
^
S = ~ YI ~D/ covariance matrix is available [30]:

F
Optimal correction Cost function ]
Cov(~}) = a2°i~eV2nnv
n~- L
1 + cos0 1
+ (sinoq/9)2 ( --
nnZ) ]
(7)
8D: , Eq.(3) Eq. (2)
where p = x/~r is the effective marker distribu-
T tion radius and r the root mean square distance
between the markers and the cluster's centroid.
Obviously, the use of Eq. 4 is more general since
no restrictions about noise and marker distribu-
tions have been introduced.
The WLS procedure is less efficient than other
E -
Estimation accuracy ] least-squares procedures, like the SVD algorithm
[10,11,13], as far as computing time requirements
} Eq. (4) I
i are concerned. It offers, nevertheless, some posi-
! tive features like the capability to adjust to noise
conditions and, above all, the possibility to
provide useful measures of the accuracy of the
DOF estimated. Moreover, its convergence is in
most cases reasonably fast, usually requiring only
Fig. 1. Flow diagram of the WLS routine, three or four iterations.
142 A. Cappello et al./ International Journal o[" Biomedical Computing 41 (1996) 137-151

2.1.1. Numerical validation

The WLS algorithm has been tested [26] in L
1.1
order to evaluate its performance in rejecting
I
stereophotogrammetric noise effects. Testing has 3 turk (2D) ]
been carried out on numerically simulated trajec-
tories of rigid clusters of three to five markers, i

fastened to a thigh during the execution of a

cycling exercise. The global coordinates of each
4rnrl~ (2D) I
marker have been altered by adding a zero-mean, 0.9
normally distributed, anisotropic noise with stan-
dard deviations o-x = O-y = 0.5 mm along the
antero-posterior axis, x, and the vertical axis, y, 4 mrl, (3D) J
0.8
respectively, and az = 1 mm along the medio-lat- ~--~,v - I ~ .~J~
eral axis, z. The WLS procedure has been applied,
using different weights for the z coordinate, and a
root mean square distance (RMSD) 0.7
0 0.25 0.5 0.75 1
Z-weight
RMSD = ~, (Yi -- ; i ) T ( y i -- Yi) (8)
i=1 Fig. 2. Influence of the weighting matrix and of the array
configuration on estimation unbiasedness.
between the original, non-noisy marker positions,
Yi, and the reconstructed ones has been evaluated
cluster during cycling) and the residuals obtained
in each case. Markers have been arranged in:
by subtracting it from the same DOF, recon-
- - an equilateral triangle with 22.3 cm sides
structed from noisy marker trajectories through a
- - a square with 15.8 cm sides
non-optimized procedure and through the WLS
- - a regular tetrahedron with 18.2 cm edges
routine. The non-optimized procedure settles a
- - a square based pyramid (base edges of 12.8
reference system on the cluster model using the
cm) with a height of 11.4 cm.
local coordinates of three of its markers and an
All these arrangements feature a quasi-constant
value of the quantity np 2 in Eq. 7, so to have analogous reference system on the actual cluster
employing the video-recorded laboratory coordi-
configurations comparable with respect to estima-
tion accuracy. nates of the same markers [31,32]. The D O F are
then reconstructed by calculating, with a simple
Fig. 2 illustrates the results obtained from the
rotation matrix product, the rotation and transla-
different test clusters by employing different z-
tion which, frame by frame, superimpose the
weights. The least biased estimate has been ob-
tained when the value 0.25 is assigned to z-weight, model reference system to the data reference sys-
tem. A reduction of the reconstruction error is
as one can be easily led to expect since noise
clearly visible when the WLS routine is used.
variance along z axis is four times as high as it is
along the two other coordinates. A gradual im-
provement is observable as the number of mark- 2.1.2. Experimental validation
ers that make up the cluster increases. Further, a The optimized D O F reconstruction routine has
visible advance in noise rejection is obtained when also been verified on experimental data. A young
switching from a plane cluster (4 marker square) male subject wearing an external fracture fixation
to a 3-D one (4 marker tetrahedron). It is quite device has been video-recorded during the execu-
important to emphasize that the height of a 3-D tion of two different motor tasks: walking on level
cluster plays a key role in unbiased D O F recon- ground and cycling. Four reflecting markers had
struction. been stuck onto the external fixture while other
Fig. 3a,b shows one of the non-noisy simulated four markers had been placed on the skin of the
sequences (relative to the above-mentioned square patient's thigh. The trajectories of all markers
 A. Cappello et al. / International Journal o f Biomedical Computing 41 (1996) 137-151 143

have been acquired with the stereophotogrammet- The estimates obtained with the optimized
ric measurement system E L I T E (BTS, Milan). No method proved to be much closer to the reference
smoothing procedures were used in this phase to DOF than those obtained with the non-optimized
avoid interference with the optimization. procedure for all the orientation components and
Since an external fixation tool provides a refer- for both motion sequences. As far as cycling is
ence which is rigid with the underlying bone, the concerned, the root mean square error (RMSE) of
anatomical segment's DOF, estimated by applying the former routine relative to the 'reference DOF'
the WLS procedure to the fixture-anchored mark- kept below 3° for all of the three components of
ers, have been assumed as being very close to the the orientation vector O, whereas the RMSE of
actual bone's DOF. Such 'reference DOF' have the latter routine was always above 3 °. Fig. 4
been compared to those reconstructed through the shows the results concerning the orientation com-
above-outlined non-optimized procedure using a ponents during walking. The advantage of the
cluster of three skin markers and to the ones WLS procedure is particularly evident on the
orientation component 0~ which shows the great-
estimated by employing the WLS routine on a
est excursion.
cluster of four skin markers.
Even though we cannot expect an evident supe-
riority of the optimized estimates for every differ-
ent motion sequence and for each possible
technical cluster, our experience is that the opti-
mized estimates are undoubtedly more accurate
on a statistical basis. Furthermore, the optimized
method sets no limits to the number of markers
which can be taken into account for DOF recon-
struction and this provides an undeniable advan-
tage over the non-optimized method, which can
2- only partially exploit the information carried by
the trajectories of three markers.
015 i 115
Time (s)
2.2. Smoothing of noisy motion data: a
a)
self-tuning digital filter

Smoothing of noisy motion data is a common

Non-optimized and critical issue whenever direct measurement of
/dll,.L .m,,!i.~i.^,A~ i~,~ [ JL li~...j~l~,~
j,j, signal derivatives is unfeasible, since differentia-
"y
,*vv,vV~"d" .... 1"*'Vw v w w ' ~ " y" ,yw,", tion magnifies the noise-to-signal ratio [33]. For
our purposes, we needed a filter which could work
Optimized without any a priori information about the opti-
• "r'vw" "I . .V,w-V- w-~ ,'r "-'V . . . . . ". . . .
mal cut-off frequency and which could, as well,
process several signals concurrently. For these
reasons, the motion reconstruction techniques we
0 015 i 1'.5 i discuss in this paper include a smoothing al-
T i m e (s) gorithm derived from the widely known 4th-order
b) digital filter obtained by using a 2nd-order Butter-
worth filter bidirectionally [34]. The main asset of
Fig. 3. (a) The simulated DOF 3,. during cycling and (b) the such a structure is that zero-phase-shift is guaran-
noise rejection performances of a non-optimized (RMSE =
6.3E-3 rad) and of the WLS (RMSE = 3.7E-3 rad) routines,
teed.
evaluated in O~ reconstruction from noisy marker trajectories The innovative feature is the algorithm's capa-
(o'noi~ = 0.5 mm). bility to self-determine the optimal cut-off fre-
144 /1. Cappello et al. / International Journal o f Biomedical Computing 41 (1996) 137 151

Thigh DOF Errors

,.-., 2 0 ~
~ 10
~ 0
~ -10

-20
0 0.2 0.4 0.6 0.8 l 0 0.2 0.4 0.6 0.8 1
20 . . . . .

~ lO
0 0

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

0k
l 0 . . . .

-30 - ' , , , ~
0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time (s) Time (s)

Fig. 4. (a) Reference o r i e n t a t i o n DOF e s t i m a t e d by f i x a t o r - e m b e d d e d technical f r a m e a n d (b) bone o r i e n t a t i o n errors between

reference values a n d those calculated from the skin technical cluster by the n o n - o p t i m i z e d (thin line) and W L S a l g o r i t h m (thick line).

quency [27,35]. This ability relies on the basic where nraf indicates that the autocorrelation
assumption that the relevant signals are affected function has been previously normalized (nraf
by additive, normally distributed, zero-mean = normalized residual autocorrelation function),
white noise. Given that, if signal and noise spectra that is divided by its own first element, the
were separate, a residual analysis performed on residual variance estimate. The theoretical opti-
the sequences obtained by subtracting the output mal value of function C should be zero. The
of an optimally tuned ideal filter from the original number of nraf elements included in the cost
noisy signals would show a total absence of auto- function represents a compromise between
correlation. The residual autocorrelation function completeness and complexity. Fig. 5a displays
(raf) would feature, in this case, a non-zero first nraf progressive reduction during cut-off fre-
element equal to the noise variance, followed by quency optimization, while Fig. 5b shows how
all zero elements. Actually, due to the filter's the cost function behaves when cut-off fre-
non-ideal configuration and to signal and noise quency is gradually increased. In fact, cut-off
frequency spectra overlapping, the above outlined frequencies lower than the optimal one produce
situation is definitely unreal. We can nevertheless high signal distortion with systematically dis-
assume that the optimal cut-off frequency is the tributed residuals. On the contrary, high cut-off
one that yields maximally uncorrelated residuals.
frequencies leave the signal unaltered but colour
Our smoothing algorithm therefore estimates the
the noise which becomes more correlated. As a
optimal cut-off frequency by minimizing the cost
matter of fact, the algorithm behaves like a
function
10
'whitening' filter which employs the autocorrela-
C = ~ nraf~i (9) tion function to test the 'whiteness' of the resid-
i=1 uals [36].
 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137-151 145

Cut-off frequency estimation is iterative: it in- 6 illustrates our filter's basic smoothing scheme
volves the assessment of the sensitivity matrix of for a single signal.
the nraf with respect to the estimated frequency When ns > 1 signals (e.g., different coordi-
and follows a pattern which parallels that of the nates or DOF) are processed simultaneously, an
DOF reconstruction procedure described in the extended version of the cost function is used
previous section. That allows to process several which takes into account the residuals of all the
signals simultaneously, thus estimating many cut- examined sequences
off frequencies at the same time. This requires ns 10
longer computing times but is sometimes neces- Cext = E Z nraJ~j (10)
sary, as in one of the filtering methods discussed i=lj=l

later, where segment DOF cut-off frequencies are The normalized residual autocorrelation func-
not estimated on the basis of DOF residual analy- tion smoothing algorithm (NRAF) has been
sis, but considering marker coordinate residuals. tested [35] on analytically defined standard refer-
In this situation, as marker positions can be only ence sequences which simulate harmonic (se-
calculated if all the segment DOF are available, quence H, [37]) and aperiodical (sequence A, [37])
the six DOF have to be filtered concurrently. Fig. kinematic signals. Twenty-five different noise se-
quences with standard deviations 0"noise 0.5 mm =

1 have been added to the simulated signals, thus

obtaining two sets of twenty five test sequences.
0.8
The percentage root mean square error [37]:
5
0.6

1 N __
0.4-
z RMSE%= ~ , ~__2~i)2_~_~s~
( x 100 (11)
0.2

-0.2
has been calculated each time to measure the
0 filtered test sequence (s) closeness of fit to the
Delay (samples) simulated signals (~). Both s and ~ include N =
a)
91 samples. The same parameter has been applied
in the comparison between the numerically as-
sessed (through a standard seven-point differenti-
ation algorithm) first and second derivatives of
the filtered test sequences and the analytically
calculated derivatives of the original signals. In all
~x cases, an established amount of sampled data (30
points) has been allocated at both ends of the
o

J Table 1
Performance of the N R A F smoothing algorithm
o
o ~ lo 1'~ 2o 2'5 3'o
Signal mean RMSE,~, _+ SD
Cut-off frequency {Hzl
b) H Displacement 1.92 _+ 0.2
Velocity 6.98 ± 0.8
Fig. 5. (a) Progressive reduction of the normalized residual Acceleration 23.1 _+ 3.5
autocorrelation function (nraf) during optimal cut-off fre- A Displacement 0.85 _+ 0.25
quency assessment and (b) cost function behavior for increas- Velocity 5.01 _+ 1.1
ing cut-off frequencies. The optimal cut-off frequency is the Acceleration 22.0 ± 5.1
one which minimizes the cost function.
146 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137-151

NRAF smoothingalgorithm

+ Illliiii_ nray . C)(nraf 'k' )

- Cost function
BW filter

i f (k+l) Cut-off t C(k)

c Frequency
Updating

Fig. 6. Flow diagram of the NRAF smoothing routine when a single signal is to be filtered.

processed sequences, in order to allow the filter to 2.3. Combined use of smoothing and optimization
reach a stationary behavior in the observation techniques: description and testing of four
interval [22]. different strategies
Test results, shown in Table 1 as mean
RMSE% over the 25 noise sequences +_ 1 stan- The combined use of smoothing and optimiza-
dard deviation, confirm that NRAF filter perfor- tion techniques can be effective in an attempt at
increasing the reliability of kinematic quantities
mances are comparable to those provided
which cannot be measured directly, but must be
by commonly used time and frequency do-
estimated starting from available noisy data.
main algorithms [22,23]. Furthermore, the tests
In this work we propose and test three different
carried out at the time when the routine was smoothing and optimization methods, later refer-
first implemented [35] showed that smoothing enced as (B), (C) and (D), comparing their perfor-
of noisy sequences H and A could be executed mances with each other and with a more
by the NRAF filter with comparable com- traditional technique, indicated with (A), that in-
putational efficiency. The above-cited remarks cludes no optimization routines. The main fea-
justify the use of the N R A F filter for our tures of the four methods are summarized in
work's aims. Table 2.

Table 2
Concise scheme of the basic features of the four methods presented and tested

Method A B C D

Filtering technique Butterworth Butterworth Butterworth Butterworth

Filtered variables Coordinates Coordinates DOF DOF
(f~ estimation (Coo. residual) (Coo. residual) (DOF residual) (Coo. residual)
criterion)
Optimization No WLS WLS WLS
procedure
Step order Smoothing+ Smoothing + Optimized DOF Optimized DOF
DOF estimate Optimized DOF estimate + estimate+
estimate smoothing smoothing
Normalized time 1.1 1.4 l 4
requirement
 A. Cappello et al. / International Journal o f Biomedical Computing 41 (1996) 137 151 147

Testing has been carried out using displace- test this fact, the comparison between the four
ment signals as well as their first and second techniques has been performed on numerically-
derivatives. All the four methods employ the simulated noisy data collected from rigid techni-
same smoothing routine, a 4th-order, zero- cal clusters.
phase-shift, self-tuning Butterworth filter, and An important difference between the tech-
the ones which include an optimization proce- niques which apply the smoothing procedure to
dure use a WLS algorithm. Both routines have coordinates, (A) and (B), and the ones which
been described earlier in this paper. apply it to DOF, (C) and (D), is that only those
The traditional technique (A) simply calcu- in the second group preserve the rigidity con-
lates the six DOF of the body segment by as- straint. After filtering marker coordinates, rela-
sessing the orientation and position vectors tive distances between markers are not constant
corresponding to the rotation and translation as they should due to the rigidity of the struc-
that make a local reference system coincide with ture they are fixed on.
a global one, determined frame-by-frame. Both The comparison among the four different
reference systems are defined starting from the techniques has been performed using simulated
spatial coordinates of three markers [31], but displacement data, which can be comparable to
while the first is relative to markers belonging to actual motion sequences. They have been ob-
the local model, the second is determined using tained through a harmonic analysis of real kine-
the actual marker global coordinates which have matic data so that their analytical expressions
been previously filtered. are always available in a simple form. This is
The second technique (B) uses, instead, the important because reference signals for the first
WLS procedure to optimally estimate the seg- and second derivatives of the displacement se-
ment DOF starting from the filtered marker co- quences, which are essential for the testing pro-
ordinates. cedure, are easily achievable by analytical
Lastly, both the third method (C) and the differentiation. The synthesis of simulated
fourth one (D) first run the WLS procedure on marker trajectories includes three main steps:
raw marker coordinates and then use the filter- Fourier analysis of segment DOF during
ing routine on the resulting DOF, but, while physical exercise
method (C) estimates the filter cut-off frequen- - Generation of simulated DOF using the
-

cies by evaluating the autocorrelation function most significant harmonic spectrum compo-
of DOF residuals, method (D) does the same by nents (in our case the first four) of the above-
considering marker coordinates residuals. This mentioned signals. Fig. 7a,b shows two
implies that in method (D) all the six segment simulated DOF.
DOF must be filtered at the same time. This - Determination of simulated marker trajecto-
-

feature makes method (D) less efficient than the ries employing a rigid plate local model and the
others as far as processing speed is concerned. DOF sequences obtained at the previous step.
The reason for considering it as an eligible tech- The cluster used in this simulation is the 15.8
nique lies in its being theoretically more correct cm sided square already employed in WLS
than method (C). The cut-off frequencies are, in testing.
fact, estimated by means of a procedure based Test sequences, later used as reference signals,
on the assumption that noise affecting the exam- have then been corrupted by superimposition of
ined signal is white. As noise is directly superim- random signals simulating additive isotropic mea-
posed to the measured marker coordinates, there surement noise (in our case relative to the
is the actual possibility that the non-linear rela- stereophotogrammetric measurement system
tionship existing between cartesian coordinates ELITE by BTS - - Milan). These signals are
and orientation vector, applied to derive the seg- pseudo-white noise sequences, normally dis-
ment DOF, colours the noise, therefore inducing tributed with zero mean and a fixed standard
a loss of efficiency in the filtering procedure. To deviation (O'noise). The last parameter determines
148 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137 151

the noise level: in accordance with A r u n et al. (2) G A I T - error sequence (a) - anoise = 1 m m
[1 l], Andrews et al. [37] and Corradini et al. [22], (3) G A I T - error sequence (b) - a,loise = 1 m m
we have considered two noise levels, setting o-,o~,e (4) C Y C L - error sequence (b) - anoise = 0.5 m m
b o t h at 0.5 m m and at 1 ram. The former value (5) C Y C L - error sequence (b) - anoi~e = 1 m m
seems to fit well the actual stereophotogrammetric (6) C Y C L - error sequence (c) - anoi~ = 1 m m
noise level, while the latter one is intended to test They have been chosen in order to produce a set
the techniques under more degraded input condi- o f results concerning an adequately wide signal
tions. sample and, in addition, to provide hints on h o w
Finally, we obtained six test sequences m a k i n g m u c h the observed performances depend u p o n the
use of: kind o f m o t i o n signal, the error sequence and the
- t w o different simulated movements concern- noise level. F o r each o f these features, there is at
ing a thigh during normal walking ( G A I T ) and least a couple o f test sequences differing just in
cycling ( C Y C L ) one relevant parameter (e.g., test sequence couples
- three error sequences: (a), (b) and (c) (2)/(3) and (5)/(6) are different merely in the error
- t w o noise levels: O'noise 0.5 m m and ano~
= sequence).
= 1 mm Technique performance evaluation has been ex-
The test sequences are patterned as follows: ecuted by c o m p a r i n g the reconstructed D O F and
(1) G A I T - error sequence (a) O-no,se = 0.5 m m their numerically evaluated (by a seven-point dif-
ferentiation routine) first and second derivatives
1.6 -] with the respective reference signals (i.e., non-
1.5- noisy test sequences and their analytical deriva-
1.4- tives). The selected quantitative parameter is the
1.3-~ well-known R M S E , a widely accepted measure o f
1.2- closeness o f fit [38]. The m o r e accurate the tech-
~ 1.1 nique, the smaller the yielded R M S E value.

0.9
0.8- 3. R e s u l t s and discussion
0.7-
0.6 As a result o f the tests performed, we have
o o15 ; 1'.5 i
T i m e (s) obtained a single R M S E value for each combina-
a) tion o f 4 techniques, 18 test sequences (6 displace-
600
ment sequences plus their first and second
derivatives) and 6 D O F .
A first survey o f this first set of data has shown
5"°1 / \ three interesting aspects:
5"°1 / \ / - - The use o f an optimization routine seems in
/ \ /
any case suitable. The R M S E values obtained
~x540 with m e t h o d (A) are the smallest only in a very
530 limited n u m b e r o f cases and these are always
520 relative to D O F that do not undergo ample
510 variations during motion.
500 / , , ,
0 0.5 1 1.5 2
- - N o n e o f the three methods, (B), (C) and
T i m e (s) (D), which include the optimization routine,
seems to be definitely outstanding. That would
b) suggest that the order in which the optimization
Fig. 7. Two simulations of thigh D O F during cycling: (a) and the s m o o t h i n g routines are run is not a
rotation D O F 3 z and (b) translation D O F T,.. crucial item. Once this hypothesis is proved on
 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137-151 149

a statistical basis, we could make a choice based Table 3

Statistical testing of the hypothesis, according to which the
on other parameters. Processing speed, for in-
four methods exhibit actual differences
stance, would indicate methods (B) and (C) as
preferable to method (D). DOF Actual ratio Actual ratio
- - Results seem to show a weak dependence of O-noise = 0.5 m m O'noise -- 1 mm

the technique performances on noise level. They 0~ 4.04 3.09

appear to be, instead, more sensitive to the test o, 0.66 0.45
reference sequence (i.e., type of movement) and oz 22.46 14.58
to the error sequence. A deeper statistical analy- Y,. 37.93 26.54
9.62
sis turned out to be worthwhile also for this T~, 19.99
Tz 1.67 2.42
purpose.
In order to statistically validate the inferences Critical value: 4.46
we have drawn from the first composite set of
data, we have repeated the performance evalua-
The figures show that we can only reject the
tion test on the simulated cycling sequence cor-
hypothesis of equivalence of the four methods if
rupted by nine different noise sequences with
we examine the three D O F which show ample
two different standard deviations (aooise = 0.5
variations during motion (i.e. #~, Tx, and T>.).
m m and ernoisc = 1 mm). Having verified that
However, if we check the test signal roughness
tests on signal derivatives agree in most cases
through the parameter
with those performed on displacement se-
quences, we have just considered the latter kind
r% = 100 O'n°ise (12)
of signals, thus obtaining two sets (each relative O-signal
to a distinct noise l e v e l ) o f six (9 noise se-
quences x 4 methods) R M S E arrays, each cor- indicated by Corradini et al. [22], we find the
responding to a single DOF. Each array row figures displayed in Table 4 (relative to test signals
contains the performance indexes (RMSE) of a corrupted with the lower level noise, anoise = 0.5
single method relative to the nine trials, while mm).
each array column includes the performance in- If we consider that L a n s h a m m a r [381 judges r%
dexes of all the four methods in a specific trial. = 0.87 as a 'realistic' value for stereophotogram-
We have subsequently examined the data metrically measured kinematic sequences and that
through a typical analysis of variance (ANOVA) Hatze [8] estimates r% = 7.28 to be relative to
strongly degraded signals, we are led to think that
technique to prove the existence of true differ-
fairly 'fuzzy' performances are plausible for the
ences a m o n g the tested methods. I f we affirm,
tested methods when random fluctuations reach
as a 'null hypothesis', that the four methods are
amplitudes as high as 20% of the signal determin-
statistically equivalent, we are allowed to reject
istic variation.
that hypothesis with a certain level of signifi-
cance in case the ratio between the variance of Table 4
the average performances of the four methods DOF roughness assessment
(variance of the array row mean values) and the
averaged performance variance (mean value of DOF r (%)
the array row variances) exceeds a certain criti- 0,. 12.60
cal value. The former quantity can be regarded Oy 23.29
as a 'justified' variance, provided the methods o~ 0.99
T~ 0.88
are actually different, while the latter expresses a T>. 0.55
strictly ' r a n d o m ' variability. Table 3 outlines the Tz 5.90
results obtained.
150 A. Cappello et al./ International Journal o f Biomedical Computing: 41 (1996) 137-151

0.35 different techniques are weakly dependent on noise

Rotation level. Method (D) proved capable of yielding the
0.3
most accurate estimates, but it is the least efficient
~0.25 when one considers computing time consumption.
8
x 0.2
4. Conclusions
,ua 0 . 1 5
(,9

,w" 0.1- The optimization and smoothing techniques pre-

sented and tested in this paper provide a reliable
0.05 -
tool for the evaluation of body segment D O F
starting from the trajectories of clusters of mark-
A B C A B C
ey
D A B C D ers. Test results, analyzed on a statistical basis,
Ox ~z
show that, if signal roughness is sufficiently low,
0.35
then the use of an optimized D O F reconstruction
Translation routine is undoubtedly advisable. This perfor-
0.3- mance has also been confirmed using stereopho-
0.25 -
togrammetric data of skin marker clusters
collected on a subject wearing an external fracture
E
0.2- fixation device which provides reference values for
~,'r" 0 . 1 5 - the bone DOF.
Smoothing of data further improves the recon-
0.1 struction accuracy while the order in which D O F
reconstruction and smoothing are carried out is
0.05
not crucial: filtering segment DOF, estimated from
0 noisy marker coordinates (methods C and D), can
B C D A B C D A B C D
Tx Ty Tz actually be as effective as reconstructing D O F from
marker coordinates which have been previously
Fig. 8. Average performances of the four tested methods. filtered (method B). Method (D) turned out to be
Accuracy in reconstructing rotation and translation DOF is
evaluated through the RMSE shown in the form mean value
the one that provides the most accurate estimates.
+ 1 std. dev. Yet, when time consumption is a priority, methods
(B) and (C) could be preferable, as they are three
Fig. 8 shows the average performances (RMSE) to four times as fast as method (D) and their
of the four methods and the relative standard performances are not far from the best obtained.
deviations.
Although restricted to displacement data, this Acknowledgements
second set of results leads to the same conclusions
as the first one: the order in which the D O F The authors wish to express their appreciation to
reconstruction and the smoothing procedure are Dr. Cesare Angeloni for his working on this
applied is not critical, but the use of an optimiza- paper's subject during the past years. His studies
tion routine is definitely advisable. This is espe- have contributed significantly to the development
cially true when the measured kinematic quantities of this work.
undergo sufficiently ample variations during the
performance of the analyzed physical exercise. The
results relative to the test signals corrupted by References
higher level noise (a,loiso = 1 mm) are in agreement [1] Andriacchi TP, Andersson JBJ, Fermier RW, Stern D
with those shown in the graphs above, thus justify- and Galante JO: A study of lower limb mechanics during
ing the hypothesis that the performances of the stair climbing. J Bone Joint Surg, 62(A) (t980) 749-757.
 A. Cappello et al. / International Journal of Biomedical Computing 41 (1996) 137 151 151

[2] Kadaba MP, Ramakrishnan HK and Wootten ME: Mea- [20] Hoppenfeld S: Physical examination of the spine and ex-
surement of lower extremity kinematics during level walk- tremities. Appleton and Lange Pub., 1976.
ing. J Orthop Res, 8 (1990) 383 392. [21] Benedetti MG, Catani F, Leardini A and Cappozzo A:
[3] Davis RB, Ounpuu S, Tyburski DJ and Gage JR: A gait Anatomical landmark definition and identification. CA-
analysis data collection and reduction technique. Hum M A R C H I n t Report, 1994.
Mov Sci, 10(5) (1991) 575 587. [22] Corradini ML, Fioretti S and Leo T: Numerical differen-
[4] Cappozzo A, Catani F, Leardini A, Benedetti MG and tiation in movement analysis: how to standardise the
Della Croce U: Position and orientation in space of bones evaluation of techniques. Med Biol Eng Comput, 31 (1993)
during movement: experimental artefacts. Clin Biomech, 11 187 197.
(1995) 90 100. [23] Gazzani F: Comparative assessment of some algorithms
[5] Cappozzo A, Catani F, Della Croce U and Leardini A: for differentiating noisy biomechanical data. Int J Biomed
Position and orientation in space of bones during move- Comput, 37 (1994) 57-76.
ment: anatomical frame determination. Clin Biomech, [24] Woltring H J: 3-D attitude representation of human joints:
10(4) (1995) 171-178. a standardization proposal. J Biomech, 27(12) (1994)
[6] Woltring H J: A Fortran package for generalized cross-val- 1399-1414.
idatory spline smoothing and differentiation. Adv Eng [25] Cappozzo A and Della Croce U: The PGD lexicon.
Software, 8 (1986) 104-113. CAMARC lI Int Report, 1994.
[7] Wood GA and Jennings LS: On the use of spline functions [26] Magnani G, Angeloni C, Leardini A and Cappello A:
for data smoothing. J Biomech, 12 (1979) 477 479. Optimal estimation of rigid body position and attitude
[8] Hatze H: The use of optimally regularized Fourier series from noisy marker coordinates. Proc X I V ISB Congress,
for estimating higher-order derivatives of noisy biome- Paris, 1993, pp. 820 821.
chanical data. J Biomech, 14 (1981) 13 18. [27] Angeloni C, Cappello A, Magnani G and Leardini A: New
[9] Spoor C and Veldpaus F: Rigid body motion calculated Weighted Least-Squares and smoothing scheme for kine-
from spatial co-ordinates of markers. J Biomeeh, 13 (1980) matics analysis: application to instantaneous helical axis
391 393. determination. Proc H I n t Symp on 3-D Analysis Hum
[10] Hanson R and Norris M: Analysis of measurements based Mov, Poitiers, 1993, pp. 57-60.
on the singular value decomposition. S l A M J Sc Stat [28] Draper NR and Smith H: Applied Regression Analysis (2''l
Comput, 2 (1981) 363 373. Ed). J Wiley and Sons, 1981.
[11] Arun K, Huang T and Blostein S: Least-squares fitting of [29] Woltring H J: 3-D 'Vectors' for joint and segment pose
two 3-D point sets. IEEE Trans Patt Anal Mach lntell, representation. Proc VIII ESB Meeting, Rome, 1992, p.
Pami-9 (1987) 698-700. 278.
[12] Veldpaus F, Woltring HJ and Dortmans L: A least-squares [30] Woltring H J: Model and measurement error influences in
algorithm for the equiform transformation from spatial data processing. In Biomechanics of human movement:
marker co-ordinates. J Biomech, 21 (1988) 45 54. applications in rehabilitation, sports and ergonomics, Bertec
[13] S6derkvist I and Wedin P/~.: Determining the movements Corporation, 1990, pp. 203 237.
of the skeleton using well-configured markers, J Biomech, [31] Chao EY and Morrey BF: Three-dimensional rotation of
12 (1993) 1473 1477. the elbow. J Biomech, 11 (1978) 57-73.
[14] Lafortune MA and Lake M J: Errors in 3-D analysis of [32] Berme N, Cappozzo A and Meglan J: Kinematics. In
human movement. Proc I I n t Symp on 3-D Analysis" of Biomechanics of human movement: applications in rehabili-
Hum Mov, Montreal, (1991) 59-62. tation, sports and ergonomics. Bertec Corporation, (1990)
[15] Karlsson D and Lundberg A: Accuracy estimation of 89 102.
kinematic data derived from bone anchored external mark- [33] Lanshammar H: On precision limits for derivatives numer-
ers. Proc II1 lnt Symp on 3-D Analysis of Hum Mov, ically calculated from noisy data. J Biomech, 15 (1982)
Stockholm, 1994, pp. 27-30. 459-470.
[16] Stanhope SJ, Holden JP and Orsini JA: Effect of target [34] Winter DA: Biomechanics and motor control of human
attachment techniques on estimates of shank skeletal movement (2 "d edtn). J Wiley and Sons, (1990) 36-41.
motion. Gait Posture, 2(1) (1994) 58. [35] Angeloni C, Leardini A and Cappello A: Smoothing of
[17] Ch6ze L, Fregly BJ and Dimnet J: A solidification proce- motion data: a new self-tuning algorithm with initial and
dure to facilitate kinematic analyses based on video system final state optimization. Proc VIII ESB Meeting, Rome,
data. J Biomech, 28(7) (1995) 879 884. 1992, pp. 277.
[18] Wang X, Rezgui MA, and Verriest JP: Using the polar [36] Ljung L: System identification: theoryfi)r the user. Prentice-
decomposition theorem to determine the rotation matrix Hall, (1987) 428.
from noisy landmark measurements in the study of human [37] Andrews BJ, Cappozzo A and Gazzani F: A quantitative
joint kinematics. Proe ll lnt Syrup on 3-D Analysis of Hum method for assessment of differentiation techniques used
Mov, Poitiers, 1993, pp. 53 56. for locomotion analysis. In Computing in medicine, The
[19] Cappello A, Leardini A, Catani F and La Palombara PF: Macmillan Press, 1982, pp. 146-154.
Selection and validation of skin array technical references [38] Lanshammar H: On practical evaluation of differentiation
based on optimal rigid model estimation. Proc I11 lnt Symp techniques for human gait analysis. J Biomech, 15 (1982)
on 3-D Analysis of Hum Mov, Stockholm, 1994, pp. 15-18. 99 105.