Professional Documents
Culture Documents
This chapter deals with differentiating in between barefoot gait and heeled gait by the means of
COP constraint curves. We assess the effect of parameters defining he COP curves’ shape factor
and bipedal model that can simulate barefoot and heeled gait. First part deals with defining the
bipedal SLIP model used in our study. The second part deals with defining the COP progression
model and its constraint curves and the algorithm to obtain steady state periodic solutions. The
third part consists of model validation by comparing the gait kinetics and spatio-temporal param-
eters for barefoot and heeled gait (Fig. 3.1). In the third part, we first test our hypothesis for
heeled gait by comparing the model’s input parameters for the best fit GRF. Then we individually
compare spatio-temporal parameters with respect to walking speed, to evaluate if our model can
distinguish between the two gaits using a singular spatio-temporal parameter. Finally, we obtain
the minimum cumulative error for heel gait determinants from the model. Based on the lowest
cumulative error ratios, we check whether the hypothesis holds true.
Figure 3.1.: Developing SLIP variant for heeled gait. Relevant gait data like physiologically
plausible TD angles, walking speed and system energy, are used to simulate the SLIP
model. Provision of COP translation is provided through a heuristic function that can
alter COP dynamics. Finally, model and experiment comparison helps to assess the
hypothetical COP curves.
To reiterate, passive dynamic walkers propagate on the principal of conservation of energy and
move under the influence of gravity. These empirical models are prescriptive and descriptive in
nature, as in their estimates can describe the human COM trajectory. Due to this they can be used
to optimize human motion. The bipedal model in this study is constrained to move in the sagittal
plane of motion. To model the equations for simulation, we employ state space representation
47
3. Estimation of heeled gait kinetics and kinematics
of the equations of motion (Appendix A.1). IP models due to their rigid legs cannot reproduce
GRF qualitatively and to suit our requirements we use the bipedal SLIP model. The model in
our study is not completely passive. The power provided by the COP movement during the first
half of stance is removed during the second half of the stance, keeping the overall system energy
constant.
mv2a k(1 − ya )2
E= + mgya + (3.1)
2 2
where, va is horizontal COM speed in the inertial frame, m is mass, g is acceleration due to gravity,
k is leg stiffness and ya is apex height (Fig. 3.2). ẏa = 0 is one of the modeling criterion in this study
which then implies va = ẋa . The energy equation is computed based on the state of the system at
apex. The apex state consists of the minimum amount of independent state variables required to
completely describe the evolution of the system over time. If the model is simulated from the
double stance phase, the state of the system would require position of COP and leg inclination
angles for both legs, apart from the COM state, that would lead to more initial conditions but
with the same outcome as the apex configuration. There exist several orbital energies E and each
energy can have infinite number of limit cycles1 . The number of limit cycles, of course, depends
upon the resolution with which the initial conditions and parameters are searched.
The leg is modeled as a linear elastic spring that stores energy during stance phase of walking.
Leg stiffness k helps in modeling the radial force in the leg during stance, based on Hooke’s law
(Eqn. 3.2) given as
Fspring = k(lo − ldim ) (3.2)
where, lo is uncompressed leg length just before heel strike and ldim is the leg length during stance.
Leg stiffness is a numerical constant that can be represented as a measure of the overall stiffness
due to joint movements at the hip, knee and ankle joint during locomotion. However, in this
study we consider the leg to be virtual connecting the COM with COP. For sake of accuracy
linear stiffness can be replaced with nonlinear stiffness as with the Duffing’s equation (k(1 −
ldim ) + η(1 − ldim )3 ) but SLIP models have shown to reproduce gait for medium to fast walking
speeds with similar GRF profile. Moreover, as shown by [66], constant leg stiffness can estimate
walking parameters at a speed of 1 m/s accurately and also at faster speeds but not at slower
speeds. This stems from the fact that the relation between GRF and compression of leg can be
modeled with a strong correlation coefficient (R2 = 0.92) using linear regression but with poor to
moderate R2 values at slower and faster speeds respectively. In this study we compare solutions
close to a walking speed of 1 m/s and thus a constant stiffness assumption suits the scope of
this study. To calculate the uncompressed leg length lo , Lipfert et al. [66] used regression to
calculate the euclidean distance between COM and COP for a given walking speed. But usually
leg length depends on the user’s anthropometric dimensions. Moreover, leg stiffness in this study
is numerically optimized due to non availability of force-length data for heeled gait.
1A limit cycle is an isolated trajectory in state space, from which he vector field either diverges outwards or inwards,
or the vector field merges into the cycle from inside or outside.
48
3.1. SLIP bipedal model
Figure 3.2.: Leg model and GRF representation. (Left) Superimposed SLIP leg with an actual
human leg. (Middle) SLIP configuration at apex state. (Right) A double stance phase
for the SLIP model, with forces experienced by the COM due to leg compression.
The force vector F acts along the direction of the leg as we consider only radial forces in the leg.
In reality due to joint torques, F acts at an angle to the leg orientation vector. But as we consider an
energy conserving model with no external torques or dissipation, GRF is directed along the virtual
leg as seen in Fig. 3.2. To obtain a generic model, we non-dimensionalize our model equations p by
dividing mass√ by m, the length terms with uncompressed leg length lo , dividing time by lo /g,
velocity by glo and acceleration by g, where g is acceleration due to gravity. The acceleration of
the COM is given as
Upon non-dimensionalization the leg force F, COM acceleration ẍ, ÿ and COM-COP inclination
angle θ during stance are defined as (Fig. 3.2)
F = k̃(1 − l) (3.3)
ẍ = FR sin θR + FL sin θL (3.4)
ÿ = FR cos θR + FL cos θL − 1 (3.5)
x − xcp
θ = tan−1 (3.6)
y
klo
p
where l = (x − xcp )2 + y2 , k̃ = mg is relative leg stiffness, k is leg stiffness, m is mass and R
and L denote right and left leg respectively. Through dimensionalization we reduce the number of
simulation parameters.
To simulate steady state of the SLIP model, we simulate three phases of walking i.e., single
stance, double stance and apex. The model alternates between single stance and double stance
phase which occur upon toe lift off (LO) and heel strike touchdown (TD) respectively. Toe LO
49
3. Estimation of heeled gait kinetics and kinematics
occurs when the F = 0 i.e., when the leg length l = 1 and the model shifts from double stance to
single stance. We use a constant TD landing angle θo . TD occurs when the vertical height of the
COM is equal to the cosine of the TD angle θo given as
yT D = cos θo (3.7)
At TD, the model shifts from single stance to double stance (Fig. 3.2). Due to these alternating
events during simulation we incorporate a switch that alternates in between 0(off) and 1(on), that
switches on when the leg is in stance and switches off when the leg lifts off. There can be various
other triggers that can be used for TD but as we restrict our steady state to two consecutive apexes,
constant TD angle condition is suitable. The swing phase of the leg does not contribute to the
dynamics of the model simulation due to the leg’s massless design. For the sake of simplicity
we search for apex states where the leg is perpendicular to the walking surface at which point
the vertical velocity of the COM is zero. In the fixed pivot SLIP, the position of the pivot on the
walking surface can be understood to be the mean position of the overall COP movement from
heel to toe occurring during the stance phase. For healthy subjects we consider the parameters
for both legs to be identical. To obtain the steady state solutions we use an error minimization
function which checks for the error in between the consecutive apex states for healthy subject
which is described ahead in the chapter.
A) B) θ1 θ2
θ1 θ2
dB dH
Figure 3.3.: COP progression for barefoot and heeled gait. θ1 corresponds to the TD angle and
θ2 is the stance angle at forefoot contact corresponding to the location of the COP. dB
and dH correspond to the COP displacement from TD to forefoot contact phase.
COP excursion is measured from the point of contact of the foot at touchdown to point of
release at toe off (Fig. 2.7). With Fig. 2.7, we show that just before mid stance, the heeled gait
will portray a larger COP displacement compared to barefoot gait. In this figure for a given TD
angle until forefoot contact phase, we hypothesize that θ1d−θ B
2
< θ1d−θ
H
2
. We believe this occurs
due to decreased stability of the heeled gait leading to faster foot rotation. COP progression
undergoes four phases of foot contact, initial contact phase, fore-foot contact phase, mid-foot
contact and toe-off [19]. They showed that COP travels around 95% of the foot length but COP
variability is observed across different gaits where the COP can travel lesser distances [78]. The
displacement limit of the COP gait line is accounted for by the foot architecture, shoe base or
physiology of the individual. COP velocity and acceleration profiles often portray high frequency
oscillations but during our modeling we estimate the filtered inherent profile of the COP, to obtain a
50
3.2. Heuristic COP model for heeled gait
regular streamlined trajectory. The COP profile for barefoot walking is obtained from the article on
bipedal walking by Jung and colleagues [49]. To overcome this variability we obtained solutions
for different magnitudes of COP displacement d during stance phase, where maximum value of d
is approximately 0.28 times the uncompressed leg length of the person, which is in essence equal
to the length of the human foot [112]. COP progression in previous studies has shown significant
movement during initial double stance for the leading leg with speeds during toe off going as
high as 22cm/s [78]. To accommodate such variability in COP displacement, we provide COP
translation throughout the stance phase compared to previous studies having COP progression
only during single stance as published by Jung and colleagues [49]. Jung’s data also shows that
COP acceleration in the fore-aft direction represents a sinusoidal curve, with acceleration in the
first half and deceleration in the second half of the stance phase (Fig. 3.4). Such a curve can
be modeled using time dependent functions. But to devise a relation between COM and COP
movement, we model the COP excursion using the independent state variables of our system.
This is achieved by obtaining a mathematical constraint relation from the hypothetical curves for
heeled gait. We obtain a constraint relation between the COP position and COM position using
a logistic function relating the COP displacement xcp and inclination angle θ (Fig. 3.4). Such a
logistic curve can be fitted with a variety of candidate functions and one such function is the tan
function that is used in our study. tan function allows for a smooth transition from barefoot COP
profile to our hypothesis constraint curve for heeled gait.
Figure 3.4.: COP constraint curves. A). Mean COP barefoot walking constraint curve (solid line)
with standard deviation (gray region) and hypothesized heeled walking COP curves
(C1, C2 and C3). B). Constraint curves (C1, C2 and C3) are a section of the tangent
function within its domain.
As mentioned above the experimental COP excursion and inclination angle relation shown in
Fig. 3.4A is approximated using a tangent function. To limit our steady state walking solutions
search, we include only those solutions for which the inclination angle at mid-stance is zero. We
assume that inclination of the leg at TD (touchdown) and LO (liftoff) is equal. At mid-stance,
we look for solutions when the COP attains half the distance of the overall COP displacement
(Fig. 3.4)A. In this way, the equation governing θ to COP excursion displayed in the Fig. 3.4A &
51
3. Estimation of heeled gait kinetics and kinematics
B is given as
θ = α tan β (3.8)
where β = (xcp − xcpo − d2 )γ, xcp is the COP position, xcpo is the COP position at TD, d is the
total displacement of the COP of an individual foot, γ and α are multiplication factors. At a given
TD angle θo , several constraint curves are generated lying in between curve C1 and C3 shown in
Fig. 3.4A. These curves simulate different COP acceleration profiles. For example, for a curve of
C1 (β ε[−π/4, π/4]), we obtain − d2 γ = −π/4 at TD, giving us γ = 2d π
. Similarly, other domains
of β are generated by changing the value of γ. Thus, by substituting Eqn. 3.6 into Eqn. 3.8 we get
x − xcp
tan−1 = α tan β . (3.9)
y
Figure 3.5.: COP translation profile reproduced from the article by Jung and colleagues [49].
A) COP displacement, B) COP velocity, C) COP acceleration
This constraint equation ensures that the logistic curve profile is enforced onto the COP dis-
placement. Eqn. 3.9 is a sufficient condition to constraint the COP. ẋcp can be greater than or
less than zero but in order to compare with the given barefoot walking profile, we include only
those constraint curves where ẋcp(T D) = 0 and ẋcp(LO) = 0. In Eqn. 3.9, value of β depends on γ
which is the shape factor of the COP displacement profile and α scales the equation such that the
magnitude θ is equal to θo at TD and LO. To obtain the acceleration curve of the COP we double
differentiate our constraint equation. Upon differentiating Eqn. 3.9 by time and rearranging, we
get
ẋ − ẋcp x − xcp 1 2
( − 2
ẏ)( (x−x 2 ) = αγ sec β ẋcp
y y cp )
1+ 2 y
˙ + yẋ − xẏ + xcp ẏ = αl 2 + αγ ẋcp sec β 2 l 2
−yxcp
ẋcp (αγl 2 sec β 2 + y) = yẋ − xẏ + xcp ẏ
52
3.3. Parameter search
Algorithm 1 Algorithm for SLIP with fixed pivot using numerical optimization
Periodic solution search for SLIP with fixed COP
for E = 1 : 0.001 : 1.3 do
for θo = 20o : 0.01 : 1o do
for k̃ = 5 : 0.1 : 200 do
for ya = q 0.9 : 0.001 : 1 do
va = 2E − 2y − k̃(1 − ya )2 , va = ẋa as ẏa = 0 at apex
Define state space equations for ẍ, ÿ
eapex = [x − xcp , y]i+1 − [x − xcp , y]i
end for
end for
end for
end for
From Alg. 1 we observe that by increments in each parameter and initial condition we can
carry out a comprehensive search to obtain a limit cycle. A stable limit cycle solution additionally
satisfies that in the neighbourhood of the fixed points the eigen values λ1,2 of the Jacobian matrix
" δ ∆x δ ∆x
#
a+1 a+1
δ ∆xa δ ya
J= δ ya+1 δ ya+1
δ ∆xa δ ya
should lie within the unit circle (|λ1,2 | < 1) [98], where the magnitude of perturbation provided is
δ ∆xa = 0.001 and δ ya = 0.001. The computational time for such a process can be decreased at
the expense of decreasing the number of periodic solutions by numerically optimizing two of the
parameters, instead of varying each parameter sequentially. For eg., in Alg. 2, by removing the
sequence for k̃, we reduce the computation time but if the return map is not convex in nature then
the optimized k̃ will iterate to solution closest to the initial guess.
Similarly, when developing a comprehensive parameter search based algorithm for SLIP with
COP translation, we carried out search for some of the relevant parameters to understand the nature
of the solutions.
With the addition of two extra parameters γ and d, the ideal scenario is to search all values of
these parameters to obtain periodic solutions as shown by the heptagon in Fig. 3.6. But we restrict
our parameter search by formulating an initial value problem which is iterated to reduce the error
in the apex return map to decrease overall computational time. To do so, we numerically optimize
[k̃, ẋcp ] and select five d = [0.05, 0.10, 0.15, 0.20, 0.25] values and three γ values of 2π 2π 2π
2d , 3d , 4d to
calculate apex states for the initial conditions specified in Table 3.1. When comparing specific data
53
3. Estimation of heeled gait kinetics and kinematics
values from heeled gait studies, in which the subjects walk at a given walking speed, our solution
search is restricted to that given speed, and searching d and γ with finer increments rather than
coarse increments to get a detailed distribution of solutions. The model is simulated in MATLAB,
Mathworks inc. 2020b using the ordinary differential equation solver package ode45 and numeri-
cal optimizer fmincon. The algorithm to obtain a closed orbit is given in Alg. 3 for the gait cycle
shown in Fig. 3.7.
In the Alg. 3, the term ∆xa+1 − ∆xa evaluates for error in the horizontal displacement of COM,
ya+1 − ya evaluates error in the vertical displacement of COM, ẋcpL(a+1) − ẋcpR(a) evaluates error
in the starting and ending position of the COP, 2(xcpL((a+1) − xcpL(T D) ) evaluates error in the total
displacement of the COP of one foot which should be equal to d, xcpL(a+1) − xa+1 evaluates error
for difference in horizontal displacement of COM with respect to COP and ẋcp checks for TD
velocity of COP.
54
3.4. Model validation and hypothesis testing
Figure 3.6.: Periodic solutions density for two scenarios of parameter search. Decrease in
solution density occurs by numerically optimizing k̃ and ẋcp .
Figure 3.7.: Gait events for a SLIP model with COP translation. SLIP model follows a se-
quence of apex, touch down (TD), lift off (LO) and apex. Such a sequence is used to
describe a periodic gait.
• Obtain periodic solutions for different γ and d values, spread across different system ener-
gies E.
• Estimate GRF curves for barefoot walking, provided by Jung and colleagues, on an absolute
time scale to check for temporal duration estimation and qualitative assessment of GRF
profiles.
• Estimate heeled gait GRF profile provided by Chien and colleagues [17] to test the heeled
gait hypothesis.
55
3. Estimation of heeled gait kinetics and kinematics
• Compare spatio-temporal estimates for heeled gait determinants to further test the hypothe-
sis.
To check if the model provides periodic solution we first simulate it using the parameters pro-
vided in Tab. 3.1.
Simulation parameter Relation Lower Bound Upper Bound
Initial Condition Energy, E 1 1.35
Apex height, ya cos θo 1
Model Parameter Leg TD angle, θo 1o 20o
2π 2π
Mult. factor, γ 4d 2.1d
COP displacement , d 0.05 0.25
Relative leg stiffness, k̃ 5 200
Table 3.1.: Scaled parameters and initial conditions used to generate periodic steady state
walking solutions. We define the COM-COP inclination angle with respect to the
vertical instead of horizontal. Multiplication factor γ is calculated as shown in Section
2.
COP translation profiles for displacement, velocity and acceleration are shown in Fig. 3.8 which
resemble those obtained experimentally shown in Fig. 3.5. Hence, our constraint function is able
to reproduce the desired curve trajectories for barefoot walking and hypothetical heeled gait COP
profiles.
Figure 3.8.: COP profile generated by the model. A) COP displacement with respect to change
in θ . B) COP velocity profile. C) COP acceleration profile. These profiles encompass
various forward progressing COP profiles including the experimental COP profiles
shown in Fig. 3.5.
With the change in γ value the model generates COP dynamics that can cater to different healthy
gait scenarios in which the inherent COP acceleration profile is sinusoidal. The sharp peak for
2π
γ = 2.1d , in Fig. 3.8B, occurs due to a slow increase in θ until close to mid-stance followed by
a rapid increase while, for other γ values ẋcp tends to increase abruptly with t. Fig. 3.8B and
2π
C is scaled with respect to the maximum value as the γ = 2.1dd had a significantly higher value
2π 2π
compared to profiles for γ = 3d and γ = 4d . The spike in COP acceleration occurs due to a larger
slope at TD and LO for ẋcp .
Distribution of stable periodic solutions corresponding to different γ and d values are shown in
Fig. 3.9. The scatter plot shows regions of isolated stable solutions which was unexpected. These
56
3.4. Model validation and hypothesis testing
Figure 3.9.: Distribution of periodic solutions at different γ and d values. Compared to SLIP
with fixed pivot, addition of two extra parameters d and γ generates additional periodic
solution plots.
isolated regions show missing blue points. The isolated regions can also correspond to bifurcation
of fixed points or the solutions converging to a local minima. For larger d values the solution
density decreases at values of E close to 1.2. As d decreases the solutions increase corresponding
to maximum k̃. The solution density is higher at larger θo and decreases as θo decreases. The
initial starting value for E = 1.35 yields no solution as we filter out solutions in the running regime
corresponding to large walking speeds.
57
3. Estimation of heeled gait kinetics and kinematics
R2 for the GRF fits, for barefoot walking, includes a time factor to take into account the temporal
error which is given as
SSE |tmodel − tdata |
R2 = (1 − )(1 − ) (3.12)
SST tdata
where SSE represents sum of squared error estimates, SST represents sum squared of total error,
tdata and tmodel are stance duration. We compare the model GRF forces for barefoot plantigrade
walking with data provided by Jung et al. [49], in which the subjects walked at 1.1m/s, 1.3 m/s
and 1.5 m/s (Fig. 3.10). At a speed of 1.1 m/s the vertical GRF shows a flatter profile around the
mid-stance which is accurately estimated by the model. The model cannot produce the heel strike
impulse observed in Fig. 3.10A which is a limitation of the simple sliding pivot of our model. The
model slightly underestimates the stance time for each case with a temporal duration error close
to 0.07s. On the contrary to our model hypothesis we obtain higher γ values for walking at 1.1
π π
m/s and 1.5 m/s which lie in the heeled gait hypothesis curve region with γ = 3.25d and γ = 4d .
Similarly, we do not observe an increasing or decreasing trend for d with increase in walking
speed. However, k̃ increases with increasing walking speed. R2 value for walking at 1.5 m/s is
slightly lower compared to the other two speeds as the model cannot estimate unequal vertical
GRF peaks due to its symmetrical modeling assumption. The issue with the model fits was having
just one experimental GRF profile instead of a collection of profiles. The increase in stiffness
at higher walking speeds is mainly due to higher energy of the system. With ya always lying in
between 0.9 and 1, with increasing E values, va and k̃ tend to increase.
A) B) C)
Figure 3.10.: GRF comparison of our model with Jung’s data. The parameters corresponding
2π
to each GRF profile are A) k̃ = 30.04, d = 0.225, γ = 3.25d , R2 = 0.82 B) k̃ = 32.07,
d = 0.125, γ = 2.25d , R = 0.83 C) k̃ = 46.43, d = 0.15, γ = 2π
2π 2 2
4d , R = 0.76.
To test our hypothesis for heeled walking, we calculate the residual error between GRF tra-
jectories and gait parameters with the data provided by Chien et al. [18]. The subjects in the
heeled gait group walked approximately at a speed of 1-1.3 m/s with heels of height hheel =
0.03, 0.05, 0.063, 0.073, 0.08m. However, the GRF was provided only for walking at heel height
of 0.073 m at a speed of 1 m/s. To compare model and experimental values we non-dimensionalize
the experimental data with similar scaling mentioned in the description of the bipedal model sec-
tion. In order to do so, we divide the distance terms by length of the leg taking into account
the increase in leg length due to the heel height (Fig. 3.11). Wearing heels lifts, increases un-
compressed leg length by elevating the COM height. The height of the leg and foot for barefoot
walking is calculated using the body part ratios provided by David winter [112]. Usage of this
anthropometric approximation leads to errors in calculation in the estimated and measured length
of the user.
Chien and colleagues provided the vertical and horizontal GRF on a %gait scale, thus, the
coefficient of determination for the GRF was calculated on a %gait scale. To test our hypothesis for
heeled gait, we employ GRF profiles corresponding to minimum single stance support%, double
58
3.4. Model validation and hypothesis testing
kleg
hH
ks
dB dH
xCP(B) xCP(H)
Figure 3.11.: Graphical representation of barefoot and heeled shoe and change in leg length
due to heel height. Heel alters effective leg stiffness of the leg and overall COP
displacement.
stance support %, cadence and stride length as shown in Fig. 3.12. The formula for error values
between model and experimental gait parameters are given by Eqn. 3.13- 3.16. To obtain % error
estimates, the error estimate for a given value is divided by the corresponding experimental value
and multiplied by 100. The model value is selected at the given walking speed corresponding to
the data and the error is estimated from the mean data point. Double stance duration is calculated
from the point when the leading leg touches down till the point when the trailing leg lifts off.
Single stance duration is calculated when the trailing leg lifts off till it again touches down. Stride
length is considered to be two times the step length. Step length is the distance between the point
where the leg lifts off till point on the floor where the same leg touches down. While cadence is
calculated as the time taken for a given step, which is then scaled to number of steps per minute.
where %e represents error, %dst represents double stance support duration percent,%sst represents
single stance support duration percent, sl represents stride length, c represents cadence. The gait
parameter values for heeled walking are (%dst )Exp. = 15.6 ± 1, (%sst )Exp. = 34.9 ± 1.4, (sl )Exp. =
1.16 ± 0.06m, (c)Exp. = 112.32 ± 9.44steps/min [18].
The GRF profile corresponding to single stance/double stance support time resembles the heeled
gait profile quite closely. On the other hand, the profile corresponding to estrln and ecad have larger
magnitudes of peak and lower magnitude of mid-stance GRF (Fig. 3.12). GRF regressed curves
show that for single stance and double stance errors, our model projects that γ = 2π 4d and d = 0.04
corresponds to this error profile, validating our hypothesis on an individual error basis. Similarly
2π
for the step length error it projects a γ = 3.5d and d = 0.03 but with a low R2vertical = 0.50 and
2
Rhorizontal = 0.23. Based on these results our hypothesis holds true for two of the gait determinants
59
3. Estimation of heeled gait kinetics and kinematics
Figure 3.12.: GRF comparison of our model with Chien’s data for heeled gait. ess,ds corre-
spond to k̃ = 85.16, d = 0.04, γ = 2π 2 2
4d , Rhorizontal = 0.84, Rvertical = 0.83, estrln corre-
2π
sponds to k̃ = 39.82, d = 0.03, γ = 3.5d , Rhorizontal = 0.50, R2vertical = 0.23 and ecad
2
2π
corresponds to k̃ = 35.62, d = 0.08, γ = 2.75d , R2horizontal = 0.48, R2vertical = 0.36.
of heeled gait. k̃ corresponding to ess GRF profile is 85.16, which equals to 41,654 N/m, for a mass
of m = 50kg and lo = 0.9972m [18]. This is a significant increase over barefoot walking k̃ at the
same walking speed. Similar to the barefoot scenario, the model cannot reproduce the heel strike
spike due to model limitations and simplistic foot model. We did not regress the entire solution
set to fit model estimates but relied on curves with least error corresponding to spatio-temporal
data points so as to simultaneously estimate kinetic and temporal predictive nature of the model.
Due to no absolute time scale representation of GRF, it was necessary to estimate spatio-temporal
errors for the model because in certain cases SLIP can underestimate temporal duration. The GRF
curve from the data were reproduced using PlotDigitizer, which is a software to extract data points
from image files.
Figure 3.13.: COP displacement d from Shang’s data plotted against a backdrop of model
parameters (k̃ and θ ) and gait events double stance duration %ds and single
stance duration %ss. For %ds, the model shows 38.24% error in predicting the
closest steady state solution at γ = 2π
4d . For %ss, the model shows 16.18% error in
predicting the closest steady state solution at γ = 2π
4d
To estimate the gait line associated with heeled gait we estimated the COP displacement d
from the study done by Shang and colleagues [91]. They showed that for heels of height 3cm,
6 cm and 8.2 cm, the gait line length (d) was around 67 cm while walking at a speed of 1 m/s.
2π
Based on this result we simulated our gait model across γ values lying in between 2.1d and 2π
4d
60
3.4. Model validation and hypothesis testing
at speed of 1 m/s, highlighting the distribution of d against k̃, θo , %ds, %ss (Fig. 3.13). For %ss
and %ds we calculate the euclidean distance error ratio for the pair (%ss, d)model and (%ss, d)data
based on which the error values in Fig. 3.13 are reported. Similar error calculation is done for
the (%ds, d)model/data pair. While estimating d and %ss the model predicts with an absolute error
% of 16.18%. And while estimating, d and %ds, the model predicts with an absolute error%
of 38.24%. Shang showed that the single and double stance corresponded to 34% of gait cycle
duration. During a gait cycle there exist two double stance phases, hence a single double stance
phase will correspond to 17% for Shang’s study. In Fig 3.13, we observe from the first subplot that
k̃ increases as the value of d decreases showing similarity to barefoot walking. Increase in d can
lead to more vertical excursion of the COM causing a compliant k̃. Similarly, θo also decreases
upon decrease in the value of d.
Figure 3.14.: Single stance duration distribution. Single stance support duration % plotted for
2π 2π 2π
γ = [ 2.1d , 3d , 4d ] for Shang and Chien’s study against model data.
To further test the validity of our hypothesis we compared the spatio-temporal values from
literature with the model predicted values of k̃, θo , ẋcp(max) , d. Our model has five parameters
in k̃, d, γ, θo , E. It is possible that when comparing model and experimental data points for a
pair of spatio-temporal parameter and walking speed, the solutions will overlap as is the case in
Fig. 3.14,3.15,3.16,3.17. In order to obtain the closest model solution to the data points plotted, we
obtain the euclidean distance between the model and data point. The % error based on these eu-
clidean distances are reported in Tab. 3.2,3.3,3.4,3.5. We observe that for both Shang and Chien’s
data, the closest model solutions show an error estimate less than 3%. These error values increase
2π
as the value of γ changes from 2.1d to 2π
4d in case of Shang’s data, while the error value is maxi-
61
3. Estimation of heeled gait kinetics and kinematics
γ k̃ θo ẋcp(max) d %error
2π
2.1d 19.62 20o 0.13 0.05 0.40
2π
Shang 3d 22.37 19o 0.03 0.05 1.75
2π
4d 22.92 18.80o 0.01 0.05 2.07
2π
2.1d 21.95 19.60o 0.14 0.05 0.27
2π
Chien 3d 26.13 17.20o 0.05 0.10 1.69
2π
4d 25.66 17.60o 0.03 0.10 0.85
Table 3.2.: Single stance duration parameters corresponding to model and data point shown
in Fig. 3.14.
Figure 3.15.: Double stance duration distribution. Double stance support duration % plotted for
2π 2π 2π
γ = [ 2.1d , 3d , 4d ] for Shang and Chien’s study against model.
mum at 2π 3d for Chien’s data. The increase in k̃ is correlated with the increase in error showing a
monotonic increase for Shang’s data points. k̃ shows correlation with error values for Chien’s data.
For both Shang and Chien the model predicts d to be less than or equal to 0.10 which proves our
hypothesis when comparing % ss. TD angle θo for Shang’s estimates show a negative correlation
with k̃ with θo being the least at γ = 2π 2π
4d and maximum at γ = 2.1d . Similar negative correlation is
observed for Chien’s model estimates. Due to lack of information on COP dynamics in either of
the studies by Shang and Chien, we would require further validation on the model predicted val-
ues on ẋcp(max) or d (in case of Chien). Double stance phase duration describes the weight transfer
phase during walking. Greater double stance duration correspond to unstable single stance gait
while lesser double stance duration corresponds to a comparatively stable gait (Fig. 3.15). Shang’s
62
3.4. Model validation and hypothesis testing
γ k̃ θo ẋcp(max) d %error
2π
γ = 2.1d 25.75 16.80o 0.15 0.05 2.78
Shang γ = 2π
3d 24.07 17.40o 0.03 0.05 3.66
γ = 2π
4d 22.92 18.80o 0.01 0.05 9.50
2π
γ = 2.1d 21.95 19.60o 0.14 0.05 2.87
Chien γ = 2π
3d 23.51 18.80o 0.03 0.05 1.07
γ = 2π
4d 72.76 10.40o 0.03 0.05 1.18
Table 3.3.: Double stance duration parameters corresponding to model and data point shown
in Fig. 3.15.
2π 2π 2π
Figure 3.16.: Cadence distribution. Cadence plotted for γ = [ 2.1d , 3d , 4d ] for Esenyel and Chien’s
study against model data.
model estimates show counter-intuitive trends compared to %ss in γ and k̃. While, k̃ and θo show
similar negative correlation as observed in %ss. The %error for Shang’s estimates for %ds is larger
compared to the %ss case with minimum error% = 2.78% and maximum error%=9.50%. While
for Chien’s case the error% are still lower with maximum error% = 2.87% and minimum error%
= 1.07%. The %ds at γ = 2π 4d shows a drastic increase in k̃ with 72.76 and significant lowering in
o
TD angle with θo = 10.40 . With slight change in error% values, k̃ and θo show large variation in
magnitudes. This warrants multiple parameter estimation which can narrow down the region for
model parameters for heeled gait.
From the model results seen in Fig. 3.16, we observe that our model predicts cadence value to
as low as 77 steps/min for a walking speed of 0.68 m/s which is an improvement compared to
63
3. Estimation of heeled gait kinetics and kinematics
γ k̃ θo ẋcp(max) d %error
2π
γ = 2.1d 26.17 19.40o 0.24 0.15 1.07
2π
Esenyel γ = 3d 31.82 20o 0.05 0.10 0.59
γ = 2π
4d 34.92 17.20o 0.04 0.20 0.38
2π
γ = 2.1d 23.67 18.80o 0.23 0.15 1.47
Chien γ = 2π
3d 28.03 18o 0.06 0.15 0.57
γ = 2π
4d 31.04 20o 0.01 0.05 1.56
Table 3.4.: Cadence distribution for parameters corresponding to model and data point
shown in Fig. 3.16.
2π 2π 2π
Figure 3.17.: Step length distribution. Step length plotted for γ = [ 2.1d , 3d , 4d ] for Esenyel and
Chien’s study against model data.
the model provided by Jung et al. [49], whose bipedal model predicts walking speeds to as low
as 0.93m/s. This plot shows that the model can also estimate the cadence of the data points from
Esenyel and Chien’s study on heeled gaits. Esenyel showed that heels of height 5.5 cm have a
cadence of 112 steps/min. Based on this data point, our model shows increase in k̃ with changing
2π
from 2.1d to 2π
4d similar to the trend for %ss and %ds for Chien’s data. COP displacement d shows
a maximum value of 0.20 and minimum value of 0.10 which is an increase from d estimates from
%ss and %ds. θo values are not correlated with k̃ for Esenyel’s model estimates. The %error is
below 1.5% for Esenyel’s data points across all γ values. For Chien’s data the model estimates do
not show a correlation between θo and k̃. The maximum %error is equal to 1.56% at γ = 2π 4d . For
step length comparison, Esenyel’s model estimates show a negative correlation between k̃ and θo
64
3.4. Model validation and hypothesis testing
γ k̃ θo ẋcp(max) d %error
2π
γ = 2.1d 25.19 19.40o 0.24 0.15 1.17
2π
Esenyel γ = 3d 33.48 17.40o 0.08 0.20 0.42
γ = 2π
4d 36.97 17.20o 0.04 0.20 0.68
2π
γ = 2.1d 25.87 20o 0.13 0.05 2.03
Chien γ = 2π
3d 26.98 17.40o 0.07 0.15 0.02
γ = 2π
4d 26.37 17.60o 0.04 0.15 0.48
Table 3.5.: Step length distribution for parameters corresponding to model and data point
shown in Fig. 3.17.
with minimum error% = 1.17%. These individual error estimates show that model estimates can
predict data points very accurately but there still exists uncertainty in testing the validity of heeled
gait hypothesis. To overcome this uncertainty we explore model capabilities in predicting multiple
parameters.
We obtain the minimum cumulative error corresponding to the heel gait determinants namely
single support% (%ss), double support% (%ds), stride length (sl ), non dimensional walking speed
(v) and stance time (st ). The cumulative error is defined as the sum of individual error ratios
corresponding to the gait determinants. To remove the bias between the error due to magnitude
of the different values of the gait determinants, we obtain the ratio of the error with respect to
the experimental value, this way we reduce the error to values lying in between 0 and 1 which
is then used to predict model values. The corresponding error values are shown in Tab. 3.6. We
2π
observe that for Chien’s study our model predicts a γ = 2.25d for barefoot walking and for heeled
2π 2π
gait it predicts γ = 3.5d , 4d which shows that the model is able to differentiate between barefoot
and heeled gait. Although, these error values project the d at 0.25 which could be explained
by the fact that our model considers the COP movement on the force plate rather than the COP
movement predicted by the insole pads inside the heel. The difference in the COP measurement is
evident as the geometry of the foot and shoe are different. This leads to different localization and
movement of pressure along the base of the foot and along the force plate. To further investigate
the difference in these two COP profiles we require more details on the movement of COP along
the force plate or along the insoles. But the model is able to predict COP displacements for lower
d values and higher d values when judging the model for both individual and multiple parameter
estimation. The error% for all cases for non dimensional velocity is less than 11% and error% in
sl is less than 11%. Error% for temporal duration are significantly low with all gait error% lying
65
3. Estimation of heeled gait kinetics and kinematics
below 3%. Moreover, the model predicts the TD angle to decrease with increasing height which
2π
would be the case for heeled gait. Similarly, the increase in k̃ with change in γ from 2.1d to 2π
4d was
expected for heeled gait.
Table 3.6.: Gait determinants error for Chien’s data. Model simulation parameters correspond-
ing to least error for gait determinants for barefoot and heeled walking.
Shang and colleagues did not provided explicit gait determinants for heeled gait and hence, we
considered the same spatio-temporal parameters as used in Chien’s study except for step length
Tab. 3.7. Instead of step length we used ddata experimental COP displacement to obtain corre-
sponding model parameters. Similar to Chien’s study the model predicts a γ = 2π 4d for Shang’s
data points. k̃ is 80.82 which is significantly large compared to Chien’s k̃ estimates. Moreover, the
model predicts close to accurate COP displacement value d with an error of 2.56%. The %error in
%ss and %ds is considerably large as compared to Chien’s model estimates. The decease in d is
accompanied by a monotonic decrease in θo for Shang’s model estimates. Power for heeled gait
Table 3.7.: Gait determinants error for Shang’s study. Model simulation parameters corre-
sponding to least error for gait determinants for heeled walking.
is calculated as
Power = F.v (3.17)
where, F is the resultant GRF and v is the resultant walking speed at a given time step tn . Heeled
gait power shows negative rebound trend and positive preload trend compared to barefoot walking.
This occurs because the model predicts a difference of less than 0.01 (BW) between the peak and
mid-stance valley of the vertical GRF for heeled gait for the dotted curve shown in Fig. 3.18.
Power profile is not known to be a gait determinant for heeled gait. Moreover, data on COM
power profile for heeled gait is required to compare the given power profile. Whereas, for barefoot
walking the model shows similar trend as observed for barefoot walking provided by Jung and
colleagues [49].
3.5. Discussion
COP dynamics
The ability of the model to generate similar COP acceleration profile shows that our model is ca-
pable in reproducing qualitatively similar COP profiles for barefoot walking as shown by Jung and
colleagues [49]. Compared to previous studies with COP translation, the constrained curves in our
66
3.5. Discussion
Figure 3.18.: Power of the COM during the four phases of leg during stance phase. Barefoot
and heeled gait curves are plotted at 1.1 m/ and 1 m/s respectively corresponding to
the GRF plots in Fig. 3.10 and Fig. 3.12. Heeled gait power curve shows opposite
behavior during rebound and pre-load stage.
model take into account the independent sate variables required to define the state of the system.
Such a description of the COP-COM dependency is analogous to the description of postural bal-
ance during quiet standing where the COM and COP movement are described using the angle of
inclination [39, 41].
The essential value of model based COP curves is that it captures rotation of the foot from TD
to LO. During TD the foot lands in the rear foot phase, consequently snapping and increasing
the xcp until apex state is achieved. We believe for heeled gait, the forefoot contact is achieved
before the leg is perpendicular to the walking surface which leads to decrease in ẍcp , attaining
zero acceleration at apex. This was also suggested by Stefanyshyn and colleagues that braking
force during heel gait is higher [97]. The experimental COP profile curves considered in our
study are an ideal scenario. In an ideal scenario, the COM decelerates and accelerates horizontally
and the COP accelerates and then decelerates. These curve represent the inherent nature of the
COP dynamics, filtering out the high frequency noise that occurs due to inter-joint movements
for healthy gait. There can be cases where the curves are not regular and symmetric about the
mid-stance. In such cases different COP progression curves can be used to obtain the overall COP
displacement but the profile of the curve has an effect on the leg stiffness and other spatio-temporal
variables as shown by our study. Compared to COP progression model proposed by Bullimore and
colleagues [11] where in they define the COP displacement to be a constant velocity model, we
67
3. Estimation of heeled gait kinetics and kinematics
use the biomechanical aspects of foot rotations to develop our hypothesis and how it could affect
the COP movement. Moreover, with the constant velocity model they observed counter-intuitive
changes in leg stiffness for human running suggesting more cautiousness towards an appropriate
choice of COP models. When choosing COP models it is important to consider the qualitative
profile of the COP progression, either in a heuristic form or the COP profile from gait data to
make the bipedal model more plausible. This was the case with Jung and colleagues [49] who
used a sinusoidal time dependent heuristic function to define COP progression which inspired
us to incorporate a non time dependent COP progression model. With our model we were able
to define COP movement based on the model’s instantaneous state. This way we are able to
reproduce COP dynamics having different shape factors that helped us generalize our model for
barefoot and heeled gait.
GRF estimation
With a generic COP progression model, the gait model estimates GRF profiles for barefoot gait
from 1.1 m/s to 1.5 m/s with strong correlation coefficients. Compared to fixed pivot SLIP, the
68
3.5. Discussion
model is able to account for temporal and spatial span observed in gait data as shown in Fig. 3.10.
However, compared to Jung’s model, it slightly underestimates the temporal factor by around
0.07s but still maintaining a strong correlation. It is able to account for qualitative GRF profile
at a speed of 1.1 m/s which could be due to the inherent SLIP model dynamics. Due to tem-
poral underestimation we observe a difference of 800 N/m in our stiffness prediction compared
to stiffness reported by Jung and colleagues. Due to the reduced parameter search, it would be
appropriate to suggest that the model skipped a given parameter set that could have provided an
accurate stance time estimate. This reduced parameter search could be the reason behind obtain-
ing underestimated stance times for barefoot GRF profiles. Additionally, the models shows an
increase in stiffness values with increasing speed also shown by other authors [34, 49]. However,
2π
the model predicts γ = 3.25d for a speed of 1.1 m/s which contradicts our hypothesis but predicts
2π
γ = 2.25d which validates our hypothesis. In summary, the barefoot gait data estimates suggest
further evidence to test the hypothesis due to lack of clarity in differentiating γ values. This am-
biguity could also be due to GRF profiles for heeled gait being similar for barefoot and heeled
gait. Due to non-distinguishable GRF profiles the data validation warranted other ways to test the
hypothesis. This way, γ values from GRF fits could not be held accountable to test the heeled gait
hypothesis.
Switching over from barefoot to heeled gait, we compared GRF profile over a %stance cycle due
to lack of absolute time data. To compensate for absolute time scale, we evaluated GRF fits based
on curves representing minimum values for %ss, %ds,%cadence and %stride length solution sets
(Fig. 3.12). This plot suggests that %ess/ds shows the best fit with R2horizontal = 0.84, R2vertical = 0.83
with k̃ = 85.16 and γ = 2π 4d . This further suggests that heeled gait and fast walking barefoot gait
possess similar γ values and could be due to the higher impact force at TD leading to a larger
denominator for γ. Compared to the smooth curvature of the human heel, heeled footwear has rigid
edges causing hard impact at heel strike making the user rotate their foot faster to obtain fore foot
contact faster. Guo et al. [36] mentioned that with heeled footwear, one observes brisk acceleration
and the high initial COP acceleration at γ = 2π 4d suggests this. Moreover, Stefanyshyn [97] also
suggested that heeled gait portrayed increase in braking force, which could be a factor of faster
foot rotation also observed in our study. Thus, trajectories lying beyond γ = γ2π den
(γden > 4d) might
provide solutions for hard sole or high heeled footwear. Usually heeled shoes with hard soles/heel
will have a rigid material with higher stiffness which is reflected by k̃ obtained for heeled gait
for Shang’s model estimates. However, we require evidence on heeled gait material stiffness to
validate the stiffness value. Barefoot walking at 1.1 m/s and heeled walking at 1.08 m/s can be
differentiated on the basis of k̃, with non heeled stiffness being 30.04. The leg length in both cases
was approximately 1m, while average mass of the subjects in Jung’s study was m = 80kg, which
makes the dimensional stiffness equal to 23,575 N/m. With a difference of 18,078 N/m between
the two walking trials, the model suggests that rigid heels could have higher stiffness.
Another important aspect of parameter estimation through bipedal models is the approximation
of dimensional values such as lo and COP distance during stance. Jung et al. [49] used a COP
translation of 0.15 m, which is approximately 60% of the foot length of the subjects in their study.
Due to the uncertainty in COP displacement observed in literature, we simulated the model for dif-
ferent COP displacements d. Grundy et al. [35] showed that the COP velocity under the metatarsal
heads can go as high as 22 cm/s and as low as 6 cm/s. This suggests that the COP progression
beyond the metatarsal phalangeal should not be discounted which is addressed in our model by
considering COP progression throughout the stance phase. Although we constraint our model to
pass zero inclination angle at mid-stance, our model can be altered to simulate asymmetric gaits
in which mid-stance and zero inclination do not occur simultaneously. This way we can simulate
gaits in individuals which deviate from the symmetrical nature of walking observed in healthy
69
3. Estimation of heeled gait kinetics and kinematics
subjects.
Multiple-parameter estimation
The model shows that for Chien’s data the estimates show very low error with heeled gait γ values
lying in between 2π 2π
3d to 4d . This proves our hypothesis to a great degree but for future research
obtaining COP displacement data for heeled gait would provide better validation opportunities.
The model does differentiate slightly in between heels of different sizes by suggesting different γ
values in Tab. 3.6. Upon that it also suggests that k̃ would increase with increase in heel height and
θo would decrease. This was also observed by Chien and colleagues showing the inclination angle
decreases with increasing heel heights [18]. As mentioned above the limiting factor is validating
the overall COP displacement which the model suggests is close to 90% of the foot length for
heeled gait. This could be due to the shape of the heel base and heel angle towards the metatarsal
joints. Similarly, for Shang’S study, the model is able to predict a very high stiffness and precise
d value with a very low θo suggesting that the model can estimate data from multiple studies
(Tab. 3.7). However, compared to Shang’s study it drastically deviates from reported %ss and
%ds. Yung Hui and Wei Hsien demonstrated increase in impact force due to increase in heel
height which is observed by the increase in leg stiffness with increase in heel heights [118]. Upon
that increase in TD angle suggests that the leg becomes steeper with increase in heel height, which
can also be extrapolated to the fact that the ankle angle would also increase corroborating the
findings by Ucanok and colleagues [102].
70
3.5. Discussion
When comparing, power phases during the stance phase, we observe that based on the mini-
mum error estimates for heeled gait, the rebound and preloading phase show opposite behavior
compared to barefoot gait (Fig. 3.18). This could be due to the rigid heel leading to decrease in
flexion of the knee causing lower rebound compared to barefoot gait, during which the muscles
flex more and provide a higher upward thrust. During, second half of stance phase, the negative
value arises from the vertically downward velocity of the COM, as both horizontal and vertical
GRF show positive values along with the horizontal COM speed. The opposite trend in heeled
gait power is also due to a flatter vertical GRF profile. Due to the plantarflexed ankle, the down-
ward movement of COM is restricted after the apex state, unlike during barefoot walking when
the ankle shifts from dorsiflexion to plantarflexion. This phase in between dorsiflexion and plan-
tarflexion can cause the walking speed in the vertical direction to increase in magnitude, which, as
our power curve suggests, is not the case in heeled gait due to minimalistic contribution from the
dorsiflexion phase. This behavior is represented in the power plot for the COM, suggesting that γ
values can explain power difference in the two categories of locomotion.
The hypothesis for heeled footwear asserts that the change in COP displacement with respect
to change in COM-COP inclination angle ∆θ is higher compared to barefoot walking. This phe-
nomenon, as expected, arises because the foot snaps faster attaining forefoot phase in heeled gait
faster compared to barefoot walking to avoid imbalance and consequently fast track the gait to-
wards stability [36]. As shown by Schwartz et al. [88] that increase in heel height causes increase
in forefoot pressure compared to rear foot showing an adapted transfer mechanism by the wearer
and further suggesting that the foot spends more time in the forefoot phase which could be due
to increase in the gradient value as suggested by our model. Kaelin et al. [50] showed that the
person adapts to the hardness/softness of the sole while walking, suggesting stiffness changes are
not only related to walking speed and cadence, but also hardness of the shoe. This is evident as,
the overall leg stiffness of the shoe might increase due to hard soles. Existing bipedal models
with or without COP excursion consider barefoot walking and the rollover shape of the foot-ankle
complex. But we provide analysis of gait using a bipedal model and COP excursion model that
can simulate heeled gait through a constraint function, thereby reducing the number of parame-
ters. Several bipedal models have been developed to estimate gait parameters and spatio-temporal
parameters for walking with a translating COP [10, 49, 53, 62, 109]. Most of these models ei-
ther have a predefined pivot architecture like the rolling foot models or an external function for
COP progression. A COP progression model dependent on COM state, like in this study, helped
us understand the inter-dependency between the COM and COP to distinguish between barefoot
walking and heeled gait. The study by Hof et al. [39] on dynamic stability, showing on how COM
controls COP movement to prevent the body from falling, is one of the inspirations for our study
in addition to the above mentioned heeled gait literature. We further took inspiration from the ex-
perimental evidence showing how COP displacement and COM-COP inclination angle are related
during barefoot walking provided by [49]. With this study we were able to develop a transition
function that could cater to barefoot and heeled gait.
θ = α tan β (3.18)
where, the shape factor γ would be different due to the flipped curves (Fig. 3.19). This will be
a simpler equation to solve compared to the previous model. However, this equation cannot be
71
3. Estimation of heeled gait kinetics and kinematics
smoothly transitioned from our hypothesized COP transition model. The goal of the hypothesis
was to provide a transition function that, based on its profile, could transition between barefoot
gait to heeled gait. This means when shifting from barefoot walking to heeled gait, the tangent
function maintains its form but with the potential model, we will need to provide a switch (discrete
value) to shift from the tangent model to this potential model. This discrete switch will lead to
a discontinuity but can be helpful in tackling other scenarios of heeled/footwear gait. Another
limitation of the study is exploring the COP constraint parameters d and γ. We believe finer mea-
surements of γ and d we could have improved the time estimates for barefoot walking. However,
with a strong correlation fit the GRF results were considered appropriate. Usually data variability
in healthy gait occurs but as we had a singular GRF profile, we could not account for other GRF
profiles. We did not explore more γ values because there was no distinguishing feature in the
periodic solution distribution plot (Fig. 3.9) and also γ beyond 2π 4d would be redundant due to no
change in shape factor. For all d and γ values the distribution was almost similar. However, in cer-
tain regions we observed isolated fixed point solutions, which could be due to unstable solutions
around that region. Bifurcation analysis to understand the nature of these points is a potential study
but was not considered relevant for this study. By obtaining branches of fixed points for stable and
unstable solutions a stability analysis can be conducted to gain insight into the stability regime of
the parameters. We also assumed the COP speed to be zero at TD and LO but literature has shown
that in many cases the starting and ending COP speed has values close to 6 to 22 cm/s [78]. By
removing the COP velocity constraint at TD and LO the error minimization functions will com-
press into eapex = [∆xa+1 − ∆xa ; ya+1 − ya ; ẋcpL(a+1) − ẋcpR(a) ; 2(xcpL((a+1) − xcpL(T D) ); xcpL(a+1) −
xa+1 ; ẋcpL(T D) −v1 ; ẋcpR(LO) −v1 ] where v1 is the COP speed at TD and LO. To obtain COP velocity
and acceleration profiles similar to barefoot walking, we imposed the TD and LO constraint for
x−x
ẋcp . As the constraint function is given as tan−1 ( y cp ) = α tan β , this is a sufficient condition to
obtain periodic solutions. However, the initial COP speed causes a discontinuity in the COP veloc-
ity profile. In that case the model could be used as a COP update model by omitting COP velocity
and COP acceleration. The error function in that case will be eapex = [∆xa+1 − ∆xa ; ya+1 − ya ].
72
3.5. Discussion
Usually COP movement depends not just on movement of plantarflexors but overall body move-
ment. Through filtering out noise from COP profiles one can generalize models for other studies
having similar COP profiles. Chiu et al. [20] showed that COP displacement profile in the AP
direction is a sheared S shape and the COP velocity peaks at mid stance. Hence, this confirms that
barefoot walking shows bell shaped curve for COP velocity. On the other hand Cornwall showed
that COP velocity is inverted bell shape which does not comply to our model assumptions [21]. In
such a case, we would require to redefine the COP movement using an alternative function of θ .
The constraint function xcp − xcpo = α2 tan θ , where α2 is the new shape factor, would be a suit-
able assumption for the inverted bell profile (Fig. 3.19). In this case the velocity would decrease
and increase towards LO. Hence, instead of relying on external functions or data, we could use
the relation θ = f (xcp ) to obtain different curves suitable for different studies. This way, COP
movement can be related to COM movement, considering passive dynamic walkers, with springy
legs. We relied on a simplified assumption that at mid-stance COP is displaced half-way and the
COM is at apex. But to involve neighbouring apex states, where this assumption does not hold
true, such as ẏa > 0 or ẏa < 0 or θa > 0 or θa < 0 would provide asymmetric GRF profiles. This
would shift the GRF mid-stance valley regions, as seen in Fig. 3.12 shift leftwards or rightwards.
Similarly, by considering the inequalities xcpa > d/2 or xcpa < d/2, we would provide a three way
asymmetrical apex state. Due to the foot architecture the COM state at TD and LO would dif-
fer and by incorporating the three-way asymmetric conditions our solutions would better explain
other heeled gait scenarios. This could be a task for future research but requires actual COP data
of heeled gait to make accurate comparisons.
That brings us to COP profile for heeled gait. Usually COP can be measured using insole, at foot
level, and at the force plate or shoe base level. These profiles capture different contact scenarios.
In order to make recommendations on relieving pressure due to heeled gait, we would require
COP profiles at both interfaces to further validate our model. That would also affect the domain
of β and α adding more complexity to the system. Issue with limit cycle optimization is that the
model does not use gait data, thus, it has to rely on several design parameters and generate several
limit cycles. With a plantarflexed ankle due to high heels, a preloaded passive spring can be used
at the SLIP pivot, instead of the COP movement. By changing the preload value, the model can
simulate heels of different height. By comparing our model and this potential model with ankle
spring, we would have a better estimate of the overall stiffness change due to heeled gait and effect
of plantarflexion on heeled gait. This would be a similar approach to the one used by Maykranz
and colleagues [69] having an ankle joint. The issue with ankle joint is shift of the COP from heel
to toe in one time step leads to an inflection point in the GRF profile.
The issue with passive dynamic walkers, especially SLIP, is simulating very slow walking gait
features. However, people wearing heels usually walk at medium walking speeds but slow walk-
ing speed representation of human walking using gait models will be a necessity for estimating
pathological gaits in the next study on post-stroke gait estimation. Due to the potential posed by
various COP models, one could conduct several studies with passive dynamic walkers to analyze
effects of potential COP constraint curves for simulating slow walking gait. By removing the hard
constraint for d, we will obtain estimates of d, which could also be an appropriate way to use the
COP model. To trigger LO, leg force should equal zero but LO can be triggered by providing an
upper limit for COP displacement d as done by Jung and colleagues [49]. In such a scenario, the
model will encounter two LO scenarios, which could be used individually or in combination, to
allow the model to select a given scenario for LO. This could provide more stance phase control
models. The stance LO trigger would be equal to LO= xcp − xcpo − d/2. Jung and colleagues
used a COP displacement limit trigger when their model transitioned from single to double stance,
however we would require data on heeled gait and COP gait line for different walking phases to
73
3. Estimation of heeled gait kinetics and kinematics
enforce similar conditions. With no feedback loop in our SLIP model, we cannot cater to fluctua-
tions in COP sway. Passive dynamic walkers are not meant to capture COP sway but could prove
beneficial when simulating heeled gaits on the verge of trips or falls. For future studies simulating
heeled gait with velocity control or hip actuation could help with understanding postural control
with heels, similar to the models developed by Sharbafi and colleagues [94]. We used only positive
COP velocities as initial condition at apex but our model can simulate negative COP velocities at
TD and LO. Such negative COP velocity scenarios could also account for COP sway negating the
need for complex sway models. This way the COP velocity bell curve will shift downwards cross-
ing into the fourth quadrant of the sagitall plane based Cartesian coordinate. Similarly, the model
could work with sinusoidal COP velocity functions as the possibilities could be endless as long as
these profiles are symmetric about vertical y-axis at mid-stance. As mentioned by Winter [111],
that the COP position is regulated through twitching of ankle plantarflexors, this twitching can be
recorded using EMG data and transformed into displacement profiles which could then be sim-
ulated using COP constraint curve. Foot drop simulators for stroke gait work towards attaining
COP movement observed in healthy gait. By using transition functions from barefoot to heeled
gait such as tan, our model can help establish a link between barefoot and heeled gait. This way
these transition functions can be used to design the floor-bed of the shoe that provides a COP
locus compensating for a subject’s pathological COP profile. Khoury et al. [52] and others [115]
showed that by providing convex shaped pads (Fig 2.6), the locus of the COP movement could
be changed. The placement of such pads on the inside of the heeled shoe could prove beneficial
in relieving foot stress and decrease excessive concentration of pressure. By using heuristic COP
functions, we could use our COP model constraint curve method to obtain locations inside of the
foot for placement of such convex pads. For example these convex pads will redirect the force F
vector experienced by the foot, consequently changing the inclination angle θ . By predetermining
the profile for F, we could simulate the SLIP to obtain a limit cycle with a constraint curve taking
into account the convex pads. To measure pressure concentration with our COP model, we can
measure the time duration and magnitude of GRF during stance. COM-COP relation should be
paramount when it comes to gait modeling using templates due to studies showing detection of
gait instabilities in elderly people by measuring the change in the COM-COP inclination angle.
Heeled gait has similarities with other foot disorder such as flat foot, which causes decrease in
plantar pressure, due to loss of heel arch [1]. As our model can simulate different COP displace-
ment values we could use our model to provide recommendation towards optimizing overall COP
displacement.
To reduce computational time, we did not exhaustively search each set of the initial conditions
that could give a unique energy value E, instead we reduced it to a numerical optimization prob-
lem, where in the stiffness and ẋcp converged to the closest minima. However, we were able to
maintain the generality of our model solutions when compared with SLIP model with fixed pivot’s
solutions. By addition of more parameters, one has to take into account an additional branch of
periodic solutions formed due to that parameter. To make fair comparisons with real world scenar-
ios of heeled gait, the model outputs should accurately represent observable kinematic and kinetic
properties. Non-observable properties such as leg stiffness can be beneficial if they can be used
towards designing orthoses or pressure relieving soft pads. Due to lack of knowledge of the elastic
properties of the rigid heel, we could not provide a fair assessment of the constant stiffness values
obtained in our study. To gain more insights into the effect of stiffness, we would require factors
such as compression of heel, excursion of COM, change in joint angles. Through more gait data,
numerical value of heel stiffness can be better measured. By obtaining force length curves from
gait data, we can apply regression to obtain stiffness parameters for the SLIP, similar to the study
by Lipfert and colleagues [66]. Chien et al. [17, 18] conducted walking trials to distinguish be-
74
3.5. Discussion
Figure 3.20.: Hill’s muscle model. Hill’s muscle model consisting of a contractile element (CE),
passive element (PE) and series element (SE) with forces F at either ends of the
muscle fibre. These muscle models are used in musculoskeletal models.
tween heeled gait and barefoot gait but similar studies should be done to obtain spatio-temporal
changes as well kinetic changes for heeled footwear to get more insights into COM and COP
relation to improve pressure relieving insoles for heels.
where k̃nl represents the non linear stiffness constant. Another stiffness model is the Hill’s equation
for the muscle, where the muscle consists of a passive element, a contractile element and an active
element (Fig. 3.20). In this case the leg can be considered as a collection of muscle fibres. Hill’s
muscle model takes into account the power generated by the muscle fibre. To obtain very slow
gait features non-linear stiffness models such as the Hill’ muscle model can prove beneficial.
However, we believe the linear stiffness model can be further tested with other COP models that
could provide stance leg control and potentially solve the slow gait phenomena limitation posed
by SLIP. The equation to Hill’s muscle model is given as
where, where F is the tension (or load) in the muscle, v is the velocity of contraction, F0 is the
maximum isometric tension (or load) generated in the muscle, a coefficient of shortening heat,
b = a · v0 /F0 and v0 is the maximum velocity, when F = 0. With this equation we could eliminate
the need to measure stiffness and the time rate of change in leg length can be associated with the
force in leg length. Through active forcing in the leg’s active element, the muscle could maintain
75
3. Estimation of heeled gait kinetics and kinematics
a certain leg length, and the error minimization functions such as for the GRF profile (Fig. 3.21)
can be used. The above equation is a realistic definition of muscular force which is usually used
in musculo-skeletal modeling but to our knowledge has not yet been used in bipedal models.
Figure 3.21.: Change in shape of the vertical GRF profile. Stance cycle optimization can be
done by using various shape profiles for vertical GRF. In the plot q represents shape
factor.
3.6. Conclusion
Walking as mentioned earlier is a combination of neuro-muscular control and traction from the
walking surface. The combination of the two dictates the overall COP movement. Hence, COP
models provide a way to amalgamate all the necessary components into a single variable which
can explain the underlying mechanism of different foot behaviour. With our heuristic COP model,
through introduction of COP shape factor γ and COP displacement d, we obtained gait determi-
nants for heeled gait. Through this study we conduct a comprehensive analysis to evaluate the
hypothesis for heeled gait. Some results suggest that increase in speed and wearing heels have the
same impact on the GRF. While, multiple parameter estimation suggests that there exists a clear
distinction between heeled gait and barefoot gait. By differentiating between heeled and barefoot
gait, we relate COM movement to COP progression which opens avenues to use our constraint
generation techniques for other gaits. The model shows ability to generate different COP dis-
placement values but we require more data on COP progression profiles for heeled gait, measured
by pressure insoles and force plates. With provision of COP update functions, various heuristic
functions can be used to analyse different footwear scenarios and consequently foot-ankle dynam-
ics.
We proposed a hypothesis for modeling heeled gait and proved it using gait kinetics and kine-
matics of heeled gait data and determinants. With this study we were able to show that the heeled
76
3.6. Conclusion
footwear properties can be converted into parameters that define the rotation of the foot during
stance phase. The rotation of the foot corresponded to several aspects such as, ankle moment, foot
orientation, ability to handle impact forces and gait stability, all of which can be represented using
a single gait parameter of γ, which is responsible in altering the sinusoidal acceleration profile of a
forward progressing COP. We were also able to show that our hypothesis holds true when account-
ing for cumulative error values corresponding to gait determinants. With this study we estimated
GRF profiles corresponding to barefoot walking along with its temporal span during stance phase.
In addition, we were able to predict power phases for heeled gait. The tangent function used to
describe the COP constraint demonstrated its versatility in handling different COP profiles corre-
sponding to different gaits. We believe such constraint curves can provide stance leg control which
could solve the very slow walking gait feature generation by the SLIP in future. In summary this
study proposes a method to test the effect of heeled gait which can be further developed to analyze
foot orthoses for pathological gait.
77