DATA SMOOTHING AND
DIFFERENTIATION PROCEDURES
IN BIOMECHANICS
Graeme A. Wood
Department of Human Movement and Recreation Studies
University of Western Australia
Nedlands, Australia
“‘When you can measure what you are speaking about, and express it in
numbers, you know something about it.’’
(William Thomson, Lord Kelvin.
Popular Lectures and Addresses, 1891-1894)
INTRODUCTION
The Problem
In many scientific fields a continuous process is measured at discrete
‘oints and, on the basis of these observations, the scientist attempts to
xplain the nature of the underlying process. These observations are, how-
ver, usually prone to error, and the scientist can be faced with the dilemma
epicted in Figure 1. .
Clearly in this example additional! observations, either at new points in t,
‘rat the same points (thereby obtaining an estimate of the magnitude of the
“he author gratefully acknowledges the valuable advice of Dr. Les Jennings and Dr.
ohn Henstridge of the Departments of Mathematics and Biometrics, The Univer-
ity of Western Australia, during the preparation of this paper.Data Smoothing and Differentiation
°o an error ?
OR
a damped oscillation ?
pe
Figure 1. Problems in fitting an empirical equation to discrete data points.
309
measurement error), would help to resolve the matter. Or, if some a priori
knowledge were available concerning the nature of the underlying function,
the problem would be simplified, but certain model parameters would still
need to be determined.
Fitting an empirical equation to data is usually done as a basis for theory
development, or as a means of obtaining new measures. Examples of the
first instance would be the identification of components in a muscular
fatigue curve (44), establishment of a force-velocity relationship (38), or
analysis of human growth curves (73), while prediction of future perform-
ance records (55) and determination of velocity and acceleration from posi-310 G.A. Wood
tion-time data would be examples of the latter instance. In all instances. sci-
entists attempt to obtain the simplest representation of the data that ade-
quately describes the underlying process, while eliminating from considera-
tion data scatter arising from experimental errors. This process is usually
referred to as ‘‘curve fitting”’ or ‘‘data smoothing.”
Measurement of interna) and external forces and the motions arising
from them is fundamental to the study of biomechanics. While in many
instances a biomechanist may only be concerned with the measurement and
description of motion, this is not an easy task, for animal and human move-
ment patterns can be very complex and the measurement process is quite
error-prone. Typically, body position is recorded at discrete points in time,
and other kinematic variables (velocity and acceleration) are obtained by
numerical differentiation. When measures of the causal variables (the mus-
cle and joint forces) or of the segmental and whole-body energetics are
required, they cannot be determined by direct means and must therefore be
derived from measures of body motion. This inverse dynamics approach
evolves from the Newtonian expressions
=iF=m-a
and
Lela
That is, the resultant force (F) and moment of force (M) acting on a body of
known mass (m) and moment of inertia (I) can be indirectly determined
from its acceleration behavior (a and a). This calculation requires not only
accurate measures of body motion, but also estimates of body segment
parameters (/5/.*
Instantaneous acceleration of a body can be produced by direct methods
through the use of electromechanical devices (accelerometers}. or, as men-
tioned above. by double differentiation of displacement-time records. The
direct approach suffers from the disadvantage that absolute body position
cannot be readily determined (62) and fixation to the body is difficult. while
numerical differentiation amplifies small errors in displacement-time data
to an extent that is alarming, as the error can occur with relatively high fre-
quency (5, /4, 49, 95).
To illustrate the differentiation problem, consider a sinusoidal motion of
frequency f as depicted in Figure 2a, with some added measurement error in
the form of another sinusoid of frequency 10 x f but of one-tenth the ampli-
tude. In communications-engineering parlance, the signal-to-noise ratio is
10:1. Upon differentiation, the noise is found to,be of equal amplitude to
the signal (Figure 2b), and when differentiated twi¢e, it becomes 100 times
the signal value (Figure 2c). That the amplitude of the first derivative is
proportional to f, and that of the second derivative is proportional to f?.
*Information on body segment parameters can be found elsewhere (61).Data Smoothing and Differentiation 311
Since the process of differentiation preferentially amplifies higher-
frequency components, the motion analyst is presented with a problem, for
human movements are generally, of low frequency /8, 95), while measure-
ment errors span the whole frequency spectrum (see Figure 2d). Therefore.
the biomechanist not only seeks an empirical function that adequately
describes the processes underlying displacement-time data, but also tries to
minimize the effect of measurement errors that will destroy the validity of
derived measures. To this end various procedures have been employed in
biomechanics. as this paper will show. While this work was written from a
biomechanist’s perspective, the material is equally applicable to any area of
the exercise and sport sciences where data smoothing/curve fitting and
differentiation/integration procedures are required.
Before proceeding, it is useful to consider the nature and sources of error,
for much of the problem lies in the manner in which experimental data are
gathered.
Sources of Error
The sources of error associated with the gathering of displacement-time
data are numerous and often mentioned by investigators, but there have
been few reported systematic studies of measurement errors. The following
list. compiled from several references (39, 66, 86, 95, 107), details some of
the potential sources of error associated with displacement-time data
obtained by the methodology most commonly used in biomechanics, i.e.,
cinematography. Error sources include misalignment of the camera; per-
spective error due to objects (subject or scale) out of the photographic
plane; stretching of film or imperfect registration of film in camera or pro-
jector; movement of camera or projector; distortion due to optical system
of camera or projector; graininess of film; movement of body segment mar-
kers in relation to joint axes of rotation, either through skin movement or
axial rotation of the segment; precision limits in the digitization process;
errors in recording temporal and spatial scales; and operator errors of
judgement and parailax in locating joint axes of rotation.
Some errors arising from these sources can be described as systematic in
that they introduce consistent biases into the data. Such errors include
image distortion, inaccurate scales, and placement of body markers. Sys-
tematic errors can also result from drift in a body marker’s position with
respect to an anatomical landmark, or from an analyst’s faulty perception
of where an anatomical point or joint center lies (these factors co-vary with
movement). Such errors can only be eliminated by adherence to sound cine-
matographic procedures /6/, 72), adoption of 3-D filming techniques (6/,
87), or adequate training of personnel. Data smoothing will seldom help in
this regard, for the apparent trends introduced into displacement-time data
through systematic errors are usually of a lower frequency than the real
motion, i.e.. less than 10 hertz (Hz).312
M