Gait recognition: highly unique dynamic plantar pressure patterns
amongst 104 individuals
Todd C. Pataky1,*, Tingting Mu2 , Kerstin Bosch3, Dieter Rosenbaum4, John Y. Goulermas5
1Department of Bioengineering, Shinshu University, Tokida 3-15-1 Ueda, Nagano, 386-8567, Japan.
2School of Computer Science, University of Manchester, Oxford Road, Manchester, M13 9PL, UK.
3Movement Analysis Lab, Social Paediatric Centre, St.-Vincenz-Hospital, Suedring 41, Coesfeld, 48653, Germany
4Movement Analysis Lab, Institute of Experimental Musculoskeletal Medicine, University Hospital Münster, Domagkstr. 3, 48149,
5Department of Electrical Engineering and Electronics, University of Liverpool, Brownlow Hill, Liverpool, L69 3GJ, UK.
Summary: 177 words (200 words max)
Media summary: 98 words (100 words max)
Main text: 6971 words (2500-8000 words)
Submitted to: Royal Society Interface
Date: 17 August 2011 (Revision #2)
Foot pressure-based biometric identification
Subject areas: Biometrics, Biomechanics, Bioengineering
Keywords: gait recognition, pedobarography, foot loading, image registration, dimensionality reduction,
SUMMARY (177 words) (200 words max)
Everyone's walking style is unique, and it has been shown that both humans and computers are very good at
recognising known gait patterns. It is therefore unsurprising that dynamic foot pressure patterns, which indirectly
reflect the accelerations of all body parts, are also unique, and that previous studies have achieved moderate-to-
high classification rates using foot pressure variables. However, these studies are limited by small sample sizes
(N<30), moderate classification rates (CR=~90%), or both. Here we show, using relatively simple image
processing and feature extraction, that dynamic foot pressures can be used to identify N=104 subjects with a CR
of 99.6%. Our key innovation was improved and automated spatial alignment which, by itself, improved CR to
over 98%, a finding that pointedly emphasises inter-subject pressure pattern uniqueness. We also found that
automated dimensionality reduction invariably improved CRs. Since dynamic pressure data are immediately
usable, with little or no preprocessing required, and since they may be collected discreetly during uninterrupted
gait using in-floor systems, foot pressure-based identification appears to have wide potential for both the security
and health industries.
MEDIA SUMMARY (98 words) (100 words max)
Everyone's walking style is unique, and camera- and force-based systems are excellent at recognising known gait
patterns. Here we show that foot pressure patterns can also be recognised, with relatively simple data processing,
with an accuracy of 99.6% (104 subjects). The key processing step was spatial alignment which, by itself,
improved accuracy to over 98%. Since foot pressure devices can be discreetly installed in the floor and are, in
general, robust to multi-subject environments because two feet cannot be in the same place at the same time, foot
pressure-based identification appears to have strong potential for security applications.
MAIN TEXT (6971 words)
(limit: 2500-8000 words)
When walking our feet interact with the ground in a stereotypical fashion: heel-strike, roll to the forefoot, then
push-off with the distal forefoot and toes  (Fig.1a). This process takes about 0.7 s when walking at normal
speeds of about 1.2 m/s. Is it possible that, within these stereotypical constraints, all individuals interact with the
Based on the gait recognition literature this seems plausible: individuals move their bodies and limbs in highly
unique and highly repeatable patterns , and camera-based computer systems can be trained to recognise these
patterns , even in adverse conditions such as poor lighting and brief exposure . We would therefore expect
these highly unique movement patterns to be reflected, to a certain extent, in our mechanical interaction with the
ground, and that computers could be similarly trained to recognise gait patterns from floor-based sensors. Indeed
floor-based gait recognition has already been highly successful. Recent examples include use of ground reaction
force (GRF) trajectories, wavelet decomposition and fuzzy set-based feature extraction to recognise individuals
with classification rates (CR) of 97%  and 99% .
While both camera-based and GRF-based gait recognition have been widely successful, both also have certain
practical limitations. Camera systems must overcome environmental noise, perspective, and other 3D calibration
problems, which state-of-the-art systems can do impressively, but with only moderate accuracy (74%) . Force
plate-systems must be quite large, at least 0.5 m long for full foot contact during non-targeted gait, but multiple
feet mustn't contact the plate at the same time, meaning that force plates cannot be positioned arbitrarily and also
that they cannot be used in multi-subject environments.
An alternative is plantar pressure imaging (PPI) . PPI systems typically consist of an array of hundreds or
thousands of pressure-sensitive sensors which are capable of characterising plantar pressure distributions at
spatial and temporal resolutions on the order of 5 mm and 100 Hz, respectively. There are a variety of PPI
technologies , but in their final form most systems are thin, flat, relatively rigid boards that can be embedded
in the floor to be flush with the walking surface. PPI systems do not suffer from environmental noise because the
foot can be very easily isolated from the environment using low-pressure thresholding. Even though an
individual may walk over a PPI plate at arbitrary angles, PPI systems also do not suffer from perspective
problems because foot images may be spatially aligned using automated registration techniques [9,10]. Finally,
high spatial and temporal resolutions mean that PPI systems can be used in multi-subject environments as all
footsteps are, by nature, spatiotemporally isolated.
PPIs are qualitatively highly unique amongst different subjects (Fig.2), and PPI-based biometric identification
has consequently also had varying degrees of success (Appendix A). Most of these studies report moderate
accuracy (80-85%), and we are aware of only four that report accuracies greater than 90% for sample sizes of at
least ten subjects:  – 98.6%,  – 96.0%,  – 93.1%,  – 92.3%. However, the maximum number of
subjects tested in these studies was eleven, and in a variety of pilot tests we were unable to reproduce the best of
these results , perhaps partly because we were unable to resolve certain ambiguities in the authors’ algorithm
descriptions. Only one study examined more than 11 subjects  (N=30), but accuracy was notably lower
(86.1%) than a previous study by the same group with fewer subjects:  – 93.1% (N=10). To date high
accuracies on samples notably larger than N=10 have only been achieved using complimentary information like
high-resolution skin prints  – 99% (N=32) or 3D foot sole shape – 98.7% (N=30), information which
cannot be readily obtained during uninterrupted gait because of lengthy scanning durations.
Of the purely PPI studies, it is notable that many have employed spatial normalisation procedures; since the foot
may adopt an arbitrary posture with respect to the PPI device, it seems logical to compensate for arbitrary
postures using spatial normalisation. However, we note that most of these studies employed decorrelation
(Appendix A), or equivalently: principal axis alignment , an approach which has been shown to yield much
poorer alignment than optimisation-based alignment procedures . It is therefore conceivable that improved
spatial alignment would yield improved biometric identification. It is also notable that previous PPI studies used
a variety of pre-selected features to be extracted from the raw data (Fig.1), but none, to our knowledge, has
conducted a systematic evaluation of the relative effectiveness of different features. The purposes of this study
were thus: (1) to explore the feasibility of PPI-based gait recognition on a larger sample of subjects (N>100), (2)
to systematically compare a variety of spatial alignment procedures, and (3) to systematically compare a variety
of features and feature extraction procedures.
Plantar pressure data were collected from 104 healthy individuals at the University of Münster (Table 1). These
data were previously used to compute a healthy 'average' pressure distribution . Data were recorded for 1.0 s
at 50 Hz using an EMED ST4 system (resolution: 5 mm) (Novel GmbH, Munich, Germany). Each subject
performed a total of ten trials of self-paced walking, five for each foot, yielding a total of 1040 3D (x, y, time)
images (Fig.1a). ‘Follow-up’ data from ten of these subjects were collected separately (Table 1);; these data were
obtained up to five years prior to the main data collection sessions. Prior to participation all subjects provided
informed consent according to the policies of the University of Münster.
The left- and right-foot images were examined separately after finding that single-foot analyses yielded
sufficiently high performance. This is justifiable, we believe, because (i) the literature shows that lower limb
dominance is poorly defined , (ii) naturally occurring gait asymmetries tend to load left and right feet
differently , and (iii) in post hoc analyses we found no systematic left-right asymmetries amongst subjects.
We may thus justifiably regard the left- and right-foot datasets as essentially independent, at least for the
purposes of validating our methods on the population from which the present subjects were drawn.
2.2 Image alignment
Images were spatially padded by adding at least 1 cm of zero pressure rows/columns to the foot periphery. They
were then temporally aligned so that the first (x,y) time slice corresponded to initial heel-strike. Following
padding all images were contained in a 65 × 29 × 50 voxel grid (x, y, time) (94,250 voxels) of which an average
of 8,291 voxels (8.8%) were non-zero for any given trial; across all subjects and trials 33,143 (35.2%) were non-
zero. The raw data were quite smooth (Fig.1a,b) so images were neither spatially nor temporally filtered.
Subsequently three categories of spatial alignment procedures were tested (Table 2). The first: 'None' performed
no alignment, passing raw images directly to feature extraction (below). The second: 'Decorrelation' performed a
principal axis transformation to centre the pressure-weighted foot centroid and to vertically align the foot's minor
principal axis. The third: 'Registration'  utilised a rapid frequency-based alignment procedure [9,23] to
automatically align a given image to a foot template. The goal of the algorithm was to maximise cross-
correlation, first in the frequency domain, to optimise horizontal and vertical foot translations, and then in the
log-polar domain, to optimise foot rotation. Example data (Fig.3) reveal that both Decorrelation and Registration
tended to improve alignment, although registration performed qualitatively better, agreeing with previous results
For Registration, seven template images were tested including: (i) the morphologically average contralateral foot
from the Münster data sample (RegMunCont) , (ii) an average foot from a separate study, a separate
laboratory, and collected with a different manufacturer's equipment , and (iii-vii) average feet of the
chronologically first five subjects from the cited study. Bilinear interpolation was used for all image
Since desktop computing memory was inadequate to submit the 3D images directly to classification routines,
and since classifiers generally perform more poorly with increasing dimensionality , the images were first
reduced to ten different 2D spatial (x,y) 'pre-features' by extracting specific characteristics of each pixel time
series (Fig.1c,d) (Table 3); we refer to these as 'pre-features' to distinguish them from the final features upon
which classification was based; the final features were extracted automatically from the pre-features using
various dimensionality reduction techniques (Sect. 2.4).
Specific pre-features included those commonly used in the plantar pressure literature: ‘peak pressure’ or
equivalently ‘maximum pressure’, or equivalently ‘100th percentile pressure’ (P100). This 2D variable represents
the maximum pressure experienced by each part of the foot over the course of stance, and is by far the most
common variable seen in the plantar pressure literature, often used to check for plantar tissue overloading .
Other common variables analysed included: the pressure-time integral (PTI), contact duration (CD), and time-to-
maximum (Tmax) . The pressure-time integral represents the total loading during stance; areas of the foot
with brief, high-pressure impulses may have a similar PTI value to areas with long, low-pressure impulses. Since
the precise variable(s) regulating plantar tissue breakdown are unknown, PTI, which quantifies loading in a
different way, has also been commonly analysed in the literature. CD is a PTI-like variable which considers only
loading duration, not magnitude, and Tmax represents yet another loading feature: loading rate (with respect to
initial heel contact). The point is that PPI data are complex, and that no single 2D variable can characterise the
3D loading profile.
In addition to these common variables, we also tested one that is less commonly used: time-to-first contact
(Tfirst)  and others that, to our knowledge, have not been previously reported the 90th, 80th, 70th, 60th, and
50th percentiles (P90, P80, P70, P60, P50). Tfirst, like Tmax, represents a specific loading-rate feature: the speed
with which one transitions to different parts of the foot. This is less common than the aforementioned variables,
most likely because load magnitude is quite low at first-contact. The percentile variables, we believed, were also
worth testing, partly because P100 is a maximum function, and therefore may be more susceptible to sensor
noise than other percentiles, and partly to check if there was a systematic effect on the ultimate results as one
considers relatively higher pressures. All aforementioned pre-features were tested either individually or in pairs,
by vectorising then stacking 2D images. Since the full image time-series were too large for practical testing, 2D
feature-pairing permitted inclusion of additional dynamic characteristics.
2.4 Dimensionality reduction
The second feature extraction phase used automated dimensionality reduction to further reduce the pre-features
to a dimensionality most effective for classification. Reduction algorithms included (Table 4): Laplacian
eigenmaps (LE) , normalised spectral clustering with a symmetric Laplacian (NCSL) , kernel principal
component analysis (KPCA) , and locally linear embedding (LLE) . Following semi-systematic analysis
(Sect. 3.1) we found that reduction to a dimensionality of 70 (from 1885 dimensions for single pre-features and
3770 dimensions for paired pre-features) worked well for these data. Other reduction parameters were manually
tuned for the left foot using 104-fold cross-validation (Sect. 2.5), and final performance was verified on the left-
foot dataset and also with a separate (leave-one-out) validation scheme. As a baseline comparison we also used
no dimensionality reduction, submitting pre-features directly to classification.
Classification of the final features was performed using nearest-neighbour (1NN) classification; this is the
simplest possible classification scheme, detecting only the image most similar to the test image (i.e. minimum
Euclidian distance) in reduced feature space. Although simple, 1NN was selected to emphasise the power of
automated dimensionality reduction for biometric-relevant feature extraction. Classifier performance was
validated using 104-fold cross-validation (104-CV) and separately using leave-one-out cross-validation to ensure
that 104-CV was not biased. We also employed a stratified 5-CV, wherein the first image of each subject was
retained for testing, while the remaining four were used for training, and then repeated for the second images,
third images, etc. This scheme (with a testing/training ratio of 25%) was adopted to ensure that the low
testing/training ratio of 0.97 % in 104-CV was not a biasing factor.
2.6 Algorithm evaluation
A full-factorial evaluation of all aforementioned factors (alignment, pre-features, dimensionality reduction
techniques, classification algorithms ) would have required a prohibitively large number of iterative tests so we
narrowed our focus by conducting semi-factorial evaluations in an ad hoc manner. For example, if variable P100
was found to perform generally better than other pre-features, then we used P100 to explore different alignment
procedures, and the resulting best alignment procedures were used to re-test all pre-features. While incomplete,
this approach proved to yield highly accurate classification performance.
Statistical hypothesis testing was conducted on a variety of classification-relevant metrics in an ad hoc manner
as context demanded. For example, a paired-sample t test was used to test whether the difference between the
None and Decorrelation alignment methods was different from zero; the motivation for this particular analysis
was to examine whether Decorrelation, the predominant alignment procedure in the literature (Appendix A), is a
better alignment choice than None. All aforementioned data processing was conducted in Matlab 7.10 (The
MathWorks, Natick, MA, USA), and all figures were created using Matplotlib 0.99 as released with the
Enthought Python Distribution 5.0 (Enthought Inc., Austin, TX, USA).
3.1 Basic results
With no image processing at all (except for image padding) nearest-neighbour classification identified
individuals with an accuracy of 90.8% using the P100 pre-feature (Fig.4). Decorrelation surprisingly yielded a
slightly lower average classification rate (CR) of 90.2%, while registration markedly increased the average CR
to 98.9%. Dimensionality reduction also tended to improve CRs (Fig.4), albeit to a lesser extent than
Across both feet the best-performing embedding dimension was 70 (Fig.5). Using this dimensionality, and
following a systematic, semi-factorial study of the different alignment algorithms, pre-features, and
dimensionality reduction schemes (Tables 5,6), the highest CR we were able to achieve in a single foot was
99.8% (519/520 correctly classified images). This was achieved on the right foot using RegMunCont alignment,
the combined P100 and P80 pre-features, and LLE dimensionality reduction. For this set of parameters the left
foot CR was 99.4% (517/520). Our semi-factorial analyses and manual parameter tuning were found to be
unbiased as leave-one-out cross validation (Table 6b), as well as validation on the left foot yielded practically
identical results (Table 6a). Additionally, we found that the low testing/training ratio of 0.97% in our validation
scheme was not a biasing factor, as a 5-CV scheme (with a testing/training ratio of 25%) yielded CRs of 99.4%
in both the left and right feet.
3.2 Follow-up dataset
Using the aforementioned 'best' parameters, CRs for the left and right feet were 98% (49/50) and 90% (45/50),
respectively, for the ten-subject follow-up dataset (Fig.6a,b). We note, however, that one of the follow-up
subjects had significantly higher right-foot metatarsal pressures in the 2007 'follow-up' trials than in the 2009
'original' trials (Fig.6c) (p=0.005, two-sample t test on extracted regional data ), and this led to 4/5
misclassifications for this subject's right foot. Upon questioning, this subject could not recall any orthopaedic
condition that could explain the 2007-to-2009 metatarsal pressure difference. We also note that all five of this
subject's left-foot follow-up images were correctly identified. If we exclude this subject’s right foot data from
follow-up analyses the CR across the nine remaining subjects would be 97.8% (44/45).
Once the classifier was trained on the 520 images from the original dataset, each follow-up image was read from
disk and classified in 2.8 and 12.5 ms, respectively, as tested on a desktop computer (2.93 GHz dual-core
processor, 4 GB memory, USB 2.0 connection to hardware) and averaged across the 100 follow-up images. Even
though data transfer delays between the pressure measurement system and PC are longer than reading from disk
(~64 ms, pilot results), a single footstep could still likely be identified within 100 ms of toe-off in a real-time
Decorrelation decreased the average CR by 3.4% and 3.6% for no-reduction and LLE-reduced data,
respectively, across all pre-features (Table 5) and both feet. After correcting for (two) multiple comparisons with
a Bonferroni threshold of p=0.025 (family-wise Type I error rate: α=0.05), paired t tests verified the significance
of this decorrelation-induced CR drop (p<0.001 and p=0.004 for None and LLE, respectively). This finding was
supported partially by root-mean-squared error (RMSE) results for the no-pre-processing, decorrelation, and
registration (RegMunCont) conditions of: 22.4±7.5, 18.0±7.2, and 12.2±5.7 kPa, respectively (mean ± st.dev.,
computed with respect to the intra-subject mean foot). It was further supported by ANOVA on no-alignment vs.
decorrelation MSE; a significant SUBJECT effect was found (p<0.001), but no significant DECORRELATION
effect was found for either the entire time series (p=0.934) or the P100 pre-feature (p=0.339). A marginal FOOT
effect was found for the time series data (p=0.070) but not for the P100 pre-feature (p=0.338); since our best-
performing classifier used only 2D pre-features (including P100) we may conclude that decorrelation's failure to
reduce intra-subject MSE was similar in both feet. . In agreement with the present CR results (Fig.4), the present
ANOVA results imply that decorrelation was not effective at reducing intra-subject variability. Therefore
choosing decorrelation over no-alignment may not be statistically justified, in general, unless initial foot posture
is highly variable. Indeed, over all tested parameter combinations registration invariably out-performed
3.4 Foot shape vs. pressure distribution
The best alignment and reduction schemes with a binary P100 pre-feature (i.e. a binary image defined by the
inequality: P100>0) yielded CRs of 93.7% and 96.5% for the left and right feet, respectively. As compared with
the continuous-pressure P100 pre-feature (Fig.1d) binary features reduced the CR by only 4.2%, suggesting that
a large proportion of the present classification-relevant information was derivable simply from 5 mm-resolution
foot shape. Nevertheless, in semi-factorial studies we were unable to achieve binary P100 performances greater
than 97%, suggesting that pressure distribution information is necessary for optimal subject identification.
The fact that essentially no processing (except for zero padding) yielded CRs greater than 90% across 104
subjects, as well as the currently best results of CR>99%, strongly suggest that PPI data contain high-quality
biometric information. This inter-subject uniqueness could only have been in embodied in plantar foot shape,
dynamic plantar pressure distribution, or both, as these constitute the only subject-specific information sources in
PPI data. The present binary image results of CR=~95%, which were very similar to previous binary image
results of CR=94.6%  clarified that foot shape itself constituted a substantial source of classification-relevant
information in the current sample. Nevertheless, the original non-binary data pushed these CRs above 99%,
suggesting that pressure patterns embody additional non-trivial inter-subject uniqueness.
In agreement with reports of high day-to-day PPI reliability , follow-up testing was also highly successful,
yielding CRs of ~98%, despite fairly extensive delays of up to five years between testing sessions. Together with
the presently estimated processing times of less than 100 ms per footstep these CR results suggest that PPI-based
biometric identification may be suitable for real-world security applications.
Recent successes in PPI-based classification of healthy foot types , pathological state , and PPI-based
fall detection  indicate that the current registration-based approach may also be useful for health-related
applications. We hope to explore some of these applications in future work.
4.2 Previous studies
The current CR results are, to our knowledge, higher than previous purely PPI-based identification studies
(Appendix A) except a previous five-subject study  (CR=100%). The best performing algorithm on a
database of at least N=10 subjects was Jung et al. : CR=98.6% (N=11), but a potential drawback to this
study was that two steps were obtained on a short (80 cm) platform; given average foot lengths of 25.5 cm 
and average stride lengths of 76 cm , subjects would have had to adopt unnaturally short strides to achieve
two complete footfalls on the measurement platform. Regardless, Jung et al.'s results imply that a larger database
of subjects may be identifiable even during unnatural or constrained gait. The remaining studies examined fewer
than twelve subjects (except for : CR=86.1%, N=32) and reported moderate CRs in the range 64-94%.
The higher current CRs can only be explained, we believe, by better data quality (spatiotemporal resolution,
accuracy, precision, etc.), better feature selection, or both. Some studies, for example, used PPI systems with
considerably less spatial resolution [12, 36, 37] (~35 mm). Others used relatively low-dimensional features like
~10-dimensional region of interest pressures  and ~100-dimensional centre of pressure (COP) trajectories
[11,15,33,39-41]; this is contrasted with the current ~8000-dimensional pre-features. Thus compression of PPI
data, either by sensor resolution or by lossy data reduction, likely sacrifices identification-relevant features.
Automated dimensionality reduction, used also in previous investigations of biomechanical (kinematic) data
[42,43], thus appears to be a more robust data compression tool.
4.3 Spatial alignment
Registration presently outperformed decorrelation over all tested parameter combinations, yielding CR
improvements on the order of 10% despite moderately high pre-registration CRs of 85% or more. Registration’s
successes are somewhat unsurprising because registration’s explicit goal is to minimise a dissimilarity metric
which, by definition, reduces intra-subject variability. Its successes are also consistent with previous reports that
a variety of registration approaches both qualitatively and quantitatively out-perform decorrelation .
It was more surprising that decorrelation performed worse than no spatial alignment in many cases. This can be
partially explained by stereotypical foot postures adopted by subjects – particularly the angle of the foot's
longitudinal axis with respect to progression direction . Decorrelation removes this information because the
foot becomes rotated to a 'vertical' posture. While registration to an arbitrary template would also remove some
of this stereotypical posture information, registration achieves better intra-subject alignment , so postural
information likely becomes less relevant once better alignment is achieved. Rather than registering to an
arbitrary template, as was done currently, it would be interesting to test a registration scheme that iteratively
registers a given PPI to a mean database image for each subject. This was not done currently because
improvements would not be noticeable beyond the present CRs of 99.6%.
As an aside we note that many previous PPI-based identification studies used decorrelation for spatial alignment
[11,15,40,45]. Despite its prevalence in previous papers, the current results strongly suggest that decorrelation is
a poor alignment choice. While we have speculated on potential mechanisms for decorrelation’s poor
performance (i.e. loss of stereotypical foot posture) it would be interesting to directly test this assertion by
incorporating initial posture as an additional feature in a decorrelated dataset. However, since we had no reason
to expect decorrelation’s poor performance prior to the present results, we leave this hypothesis for future work.
We wish to emphasise that we do not believe that the current registration scheme  was particularly special in
terms of generating higher CR; there are a plethora of registration algorithms in the literature , and indeed a
variety of methods have been shown to yield similar results in plantar pressure data . Furthermore, in post
hoc analyses we employed a completely different registration scheme  and achieved similarly high, albeit
slightly lower CRs of ~97.5%. The current algorithm was selected simply because it was fast and has worked
well recently. To rule out a particular registration scheme as a limitation it would be prudent to evaluate other
algorithms in future work.
4.4 Feature extraction
The best-performing single pre-features were P100, P90, P80 and PTI (Table 5), and the best pre-feature
combination of P100,P80 only marginally improved ultimate CRs (Table 6). This gives anecdotal credence to
the extensive use of P100 and PTI in the literature  as information-dense parameters. To our knowledge P90
and P80 have not been previously examined. One explanation for the success of the P100,P80 combination is
that this essentially represents a dynamic gradient, albeit a low-frequency one, and that this low-feature gradient
also contains subject-specific information. However it does not explain why P100&P80 was better than
P100&P90. Regardless, since the performances of the P80, P90, and P100 pre-features were all quite high, a
systematic exploration of their differences would not be possible without more data.
Moreso than particular pre-feature selections, and with the exception of KPCA, dimensionality reduction was
found to invariably improve CR (Table 6), albeit to a smaller extent than registration. While the CR
improvement was small it was nontrivial, pushing the average CR beyond what was achievable with raw-
spatially aligned pre-features. We may thus conclude that while certain pre-features perform very well, only with
dimensionality reduction can optimum CR be achieved. In other words, there are classification-relevant patterns
in the pre-features that cannot be extracted in an a priori manner.
As an aside we note that the present percentile pre-features (P90, P80, …) were computed over all time frames
(Fig.1c), and are therefore dependent on both the duration of supra-zero pressure and the recording duration (1
s). In post hoc analysis we also computed percentiles over contact duration, but we found little qualitative effect
on the ultimate results: P100 was the best-performing percentile, and CR systematically reduced with percentile
We also wish to restate that we presently did not conduct temporal normalisation (aside from heel-strike
alignment). This was done deliberately, to give time-related features like contact duration (CD) and time-to-max
(Tmax) the maximum chance find temporal differences amongst subjects; if there were indeed significant
temporal differences amongst subjects these features would be expected to yield higher CRs than if the data were
temporally normalised. However, the fact that CD and Tmax performed relatively poorly (Table 5) suggests that
inter-subject temporal differences were not as important as the pressure-related differences.
Finally, the present pre-feature list was incomplete. All 2D (x,y) pre-features the data were derived from the
original 3D (x,y,time) image, but additional variables could have been analysed like the spatial pressure gradient
 and the spatiotemporal (x,time) 100th percentile . It may be informative to investigate such variables in
A major practical limitation of the current study is that we investigated only unshod walking. It is conceivable
that shod walking considerably distorts classification-relevant pressure patterns and/or that subjects are not
recognisable if they wear different shoes. A second key limitation of this study is that only natural self-paced
walking data were collected; PPI data are known to change with walking speed , fatigue , and a variety
of other factors , and we note that some previous PPI-based identification studies have indeed incorporated
some of these factors in experimental classification tests .
Walking speed, in particular, would be interesting to consider; although general foot morphology does not
change with speed, and thus binary features (Section 3.4) should be largely unaffected, the non-trivial pressure
redistributions associated with walking speed  would likely affect subject separability, and it would be
prudent to empirically define the walking speed limits that retain separability. However, since we can easily
measure walking speed using cameras, and/or using foot-contact duration as a proxy, it may be possible to
algorithmically compensate for walking speed variability, for example by introducing temporal normalisation, or
by scaling pressures in certain foot regions.
Although PPI data can easily change, many gait-recognition applications involve desired identification,
situations in which an individual wants to be identified (e.g. automated airport immigration control). For other
applications it may be necessary to test the current algorithms on experimentally manipulated gait. Finally, we
presently considered only particular testing/training ratios in our model assessment. It would be prudent to
systematically explore testing/training ratios, with more images for each subject, to find the optimum number of
images one should obtain if implementing a real-world plantar pressure-based identification scheme.
Normal self-paced unshod walking produced a high-quality plantar pressure-derived biometric, and the present
identification implementation yielded classification rates of 99.6% in N=104 individuals. These results were
largely driven by spatial image registration and, to enable finer subject differentiation, automated dimensionality
reduction. Since plantar pressure data are highly unique amongst individuals, and since data can be easily
collected and processed using commercial in-floor hardware, plantar pressure-based identification appears to
have strong potential for a variety of security and health applications.
Funding for this work was provided by Special Coordination Funds from MEXT, Japan.
Blanc, Y., Balmer, C., Landis, T., & Vingerhoets, F. 1999 Temporal parameters and patterns of the foot roll
over during walking: normative data for healthy adults. Gait Posture 10(2), 97-108.
Cutting, J. E. & Kozlowski, L. T. 1977 Recognizing friends by their walk: gait perception without
familiarity cues. Bull Psychon Soc 9(5), 353-356.
Nixon, M. S. & Carter, J. N. 2006 Automatic recognition by gait. Proc IEEE 94, 2013-2024.
Stevenage, S. V., Nixon, M. S. & Vince, K. 1999 Visual analysis of gait as a cue to identity. Appl Cogn
Psychol 13(6), 513-526.
Yao, Z.-M., Zhou, X., Lin, E.-D., Xu, S. & Sun Y.-N. 2010 A novel biometric recognition system based on
ground reaction force measurements of continuous gait. In Proceedings of the IEEE International
Conference on Human System Interactions, pp. 452-458.
Moustakidis, S. P., Theocharis, J. B. & Giakas, G. 2009 Feature extraction based on a fuzzy complementary
criterion for gait recognition using GRF signals. In Proceedings of the Mediterranean Conference on
Control and Automation, pp. 1456-1461
Goffredo, M., Bouchrika, I., Carter, J.N. & Nixon, M.S. 2010 Self-calibrating view-invariant gait
biometrics. IEEE Trans Syst Man Cybern B Cybern 40, 997-1008.
Rosenbaum, D. & Becker, H. P. 1997 Plantar pressure distribution measurements: technical background and
clinical applications. Foot Ankle Surg 3(1), 1-14.
Facey, O. E., Hannah, I. D. & Rosen, D. 1993 Analysis of the reproducibility and individuality of dynamic
pedobarograph images. J Med Eng Technol 17(1), 9-15.
10 Pataky, T. C., Goulermas, J. Y. & Crompton, R. H. 2008 A comparison of seven methods of within-subjects
rigid-body pedobarographic image registration. J Biomech 41(14), 3085-3089.
11 Jung, J.-W., Bien, Z. & Sato, T. 2004 Person recognition method using sequential walking footprints via
overlapped foot shape and center-of-pressure trajectory. IEICE Transactions on Fundamentals of
Electronics, Communications and Computer Sciences E87-A(6), 1393-1400.
12 Yun, J., Woo, W. & Ryu, J. 2005 User identification using user’s walking pattern over the ubiFloorII. In
Lecture Notes in Computer Science: Computational Intelligence and Security 3801, 949-956.
13 Yamakawa, T., Taniguchi, K., Asari, K., Kobashi, S. & Hata, Y. 2008 Biometric personal identification
based on gait pattern using both feet pressure change. In Proceedings of the IEEE World Automation
14 Qian, G., Zhang, J. & Kidane, A. 2010 People identification using floor pressure sensing and analysis. IEEE
Sens J 10(9), 1447-1460.
15 Takeda, T., Taniguchi, K., Asari, K., Kuramoto, K., Kobashi, S. & Hata, Y. 2009 Biometric personal
authentication by one step foot pressure distribution change by load distribution sensor. In Proceedings of
the IEEE International Conference on Fuzzy Systems, pp. 906-910.
16 Uhl, A. & Wild, P. 2008 Footprint-based biometric verification. J Electron Imaging 17(1), 011016.
17 Hild, M. 2008 Person identification based on barefoot 3D sole shape. In Lecture Notes in Computer Science:
Computational Forensics (ed. Srihari, S. N. & Franke, K.) 5158, 84-95.
18 Harrison, A. J. & Hillard, P. J. 2000 A moment-based technique for the automatic spatial alignment of
plantar pressure data. Proceedings of the Institute of Mechanical Engineers, Part H: Journal of Engineering
in Medicine 214(3), 257–264.
19 Pataky, T. C., Bosch, K., Mu, T., Keijsers, N. L. W., Segers, V., Rosenbaum, D. & Goulermas, J. Y. 2011
An anatomically unbiased foot template for inter-subject plantar pressure evaluation. Gait Posture. 33(3):
20 Sadeghi, H., Allard, P., Prince, F. & Labelle, H. 2000 Symmetry and limb dominance in able-bodied gait: a
review. Gait Posture 12, 34-45.
21 Herzog, W., Nigg, B. M., Read, L. J. & Olsson, E. 1989 Asymmetries in ground reaction force patterns in
normal human gait. Med Sci Sports Exerc 21(1), 110-114.
22 Maintz, J. B. A. & Viergever, M. A. 1998 A survey of medical image registration. Med Image Anal 2(1), 1-
23 Oliveira, F. P. M., Pataky, T. C. & Tavares, J. M. R. S. 2010 Registration of pedobarographic image data in
frequency domain. Computer Methods in Biomechanics and Biomedical Engineering 13(6), 731-740.
24 Raudys, S. & Pikelis, V. 1980 On dimensionality, sample size, classification error and complexity of
classification agorithm in pattern recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence 2(3), 242-252.
25 De Cock, A., De Clercq, D., Willems, T. & Witvrouw, E. 2005 Temporal characteristics of foot roll-over
during barefoot jogging: reference data for young adults. Gait Posture 21(4), 432-439.
26 Belkin, M. & Niyogi, P. 2002 Laplacian eigenmaps and spectral techniques for embedding and clustering.
In Advances in neural information processing systems vol. 14, (ed. T. G. Dietterich, S. Becker, & Z.
Ghahramani), pp. 585-591. Cambridge, MA: MIT Press.
27 Schölkopf, B., Smola, A. J. & Müller, K. R. 1998 Nonlinear component analysis as a kernel eigenvalue
problem. Neural Comput 10(5), 1299-1319.
28 Roweis, S. T. & Saul, L. K. 2000 Nonlinear dimensionality reduction by locally linear embedding. Science
29 Gurney, J. K., Kersting, U. G. & Rosenbaum, D. 2008 Between-day reliability of repeated plantar pressure
distribution measurements in a normal population. Gait Posture 27(4), 706-709.
30 De Cock, A., Willems, T., Witvrouw E., Vanrenterghem J. & De Clercq, D. 2006 A functional foot type
classification with cluster analysis based on plantar pressure distribution during jogging. Gait Posture 23(3),
31 Jeon, H.-S., Han, J., Yi, W.-J., Jeon, B. S. & Park, K. W. 2008 Classification of Parkinson gait and normal
gait using spatial-temporal image of plantar pressure. In Proceedings of the IEEE Engineering in Medicine
and Biology Society Conference, pp. 4672-4675.
32 Sazonov, E. S., Bumpus, T., Zeigler, S. & Marocco, S. 2005 Classification of plantar pressure and heel
acceleration patterns using neural networks. In Proceedings of the IEEE International Joint Conference on
Neural Networks, pp. 3007-3010.
33 Jung, J.-W., Bien, Z., Lee, S.-W. & Sato, T. 2003 Dynamic footprint based person identification using mat-
type pressure sensor. In Proceedings of the IEEE Engineering in Medicine and Biology Society Conference,
pp. 2937- 2940.
34 Xiong, S. Goonetilleke, R. S., Witana, C. P. & Lee Au Y. 2008. Modelling foot height and foot shape-
related dimensions. Ergonomics 51(8), 1272-1289.
35 Ostrosky, K. M., Van Swearingen, J. M., Burdett, R. G. & Gee, Z. 1994. A comparison of gait
characteristics in young and old subjects. Physical Therapy 74(7), 637-644.
36 Yun, J., Abowd, G., Woo, W. & Ryu, J. 2007 Biometric user identification with dynamic footprint. In
Proceedings of the IEEE International Conference on Bio-Inspired Computing: Theories and Applications,
37 Yun, J., Abowd, G. D. & Ryu, J. 2008 User identification with user's stepping pattern over the UbiFloorII.
Int J Pattern Recognit 22(3), 497-514.
38 Yang, Y.-T., Lin, Y.-C. & Yang, B.-S. 2010 Gait recognition by features of plantar pressure. In Proceedings
of the World Congress of Biomechanics, 243.
39 Jung, J.-W., Sato, T. & Bien, Z. 2002 Unconstrained person recognition method using dynamic footprint. In
Proceedings of the 18th Hungarian-Korean Seminar, pp. 129-137.
40 Jung, J.-W., Sato, T. & Bien, Z. 2004 Dynamic footprint based person recognition using a hidden Markov
model and a neural network. Int J Intell Syst 19(11), 1127-1141.
41 Qian, G., Zhang, J. & Kidane, A. 2008 People identification using gait via floor pressure sensing and
analysis. In Lecture Notes in Computer Science: Smart Sensing and Context (ed. Roggen, D., Lombriser, C.,
Tröster, G., Kortuem, G. & Havinga, P.) 5279, 83-98.
42 Goulermas, J. Y., Findlow, A. H., Nester, C. J., Howard, D. & Bowker, P. 2005 Automated design of robust
discriminant analysis classifier for foot pressure lesions using kinematic data. IEEE Trans Biomed Eng
43 Mu, T., Pataky, T. C., Findlow, A. H., Aung, M. S. H. & Goulermas, J. Y. 2010 Automated nonlinear
feature generation and classification of foot pressure lesions. IEEE Trans Inf Technol Biomed 14(2), 418-
44 Taranto, J., Taranto, M. J., Bryant, A. & Singer, K. P. 2005 Angle of gait: a comparative reliability study
using footprints and the EMED-SF®. The Foot 15(1), 7-13.
45 Nakajima, K., Mizukami, Y., Tanaka, K. & Tamura, T. 2000 Footprint-based personal recognition. IEEE
Trans Biomed Eng 47(11), 1534-1537.
46 Prabhu, K. G., Patil, K. M. & Srinivasan, S. 2001 Diabetic feet at risk: a new method of analysis of walking
foot pressure images at different levels of neuropathy for early detection of plantar ulcers. Med Biol Eng
Comput 39(3), 288-293.
47 Rosenbaum, D., Hautmann, S., Gold, M. & Claes, L. 1994 Effects of walking speed on plantar pressure
patterns and hindfoot angular motion. Gait Posture 2(3), 191-197.
48 Stolwijk, N. M., Duysens, J., Louwerens, J. W. K. & Keijsers, N. L. W. 2010 Plantar pressure changes after
long distance walking. Med Science in Sports Exerc 42(12), 2264-2272.
Table 1. Subject characteristics.
3.2 a (2.2)
(Averages, with st.dev. in parentheses. 'Follow-up' data included five females and five males; main dataset: Spring 2009, two follow-up
subjects: 1.5 years later, eight follow-up subjects: 1.5-5.0 years before. aData for two female subjects were unavailable.)
Table 2. Spatial alignment methods.
No pre-processing (only image cropping
Decorrelation (principal axis
Register to (flipped) Münster mean
Register to independent ipsilateral mean
Register to independent arbitrary subject 1
Register to independent arbitrary subject 2
Register to independent arbitrary subject 3
Register to independent arbitrary subject 4
Register to independent arbitrary subject 5
Table 3. Pre-feature descriptions.
Pre-feature Units Description Definition
P100 kPa 100th percentile (spatial maximum)
P?(?,?) ≡ ?(?,?,?) +
where ??≤ ? ≤ ??+1
100(? − ??)[?(?,?,? + 1) − ?(?,?,?)]
PTI kPa·s Pressure-time integral
???(?,?) = ??(?,?,?)
Tfirst s Time to first contact (from heel strike)
??????(?,?) = min
Tmax s Time to maximum (from heel strike)
????(?,?) = argmax
CD s Contact duration
??(?,?) = ??(?,?,?)
where ?(?,?,?) = ?1 if?(?,?,?) > ?
Here I(x,y,t) is the image time series, p denotes percentile, k indexes the ordered observations of a particular pixel’s time series, pk is the
percentile of the kth ranked observation, N is the number of observations, and ε is a pressure threshold (manufacturer-set to ε =5 kPa in
the current dataset) ). Note: in the percentile equation k is not a time index, but rather indexes sorted observations, and k may be different
for each pixel’s time series.
Table 4. Dimensionality reduction methods.
None No reduction
LE Laplacian eigenmaps
Normalised spectral clustering
with symmetric Laplacian
Kernel principal component
LLE Locally linear embedding
Table 5. Semi-factorial analysis: alignment and single pre-features, left foot.
None 94.6 94.8
Decorr 93.1 93.7
RegMunCont 98.8 98.8
RegIndepMean 92.9 93.3
RegIndep1 99.0 99.2
RegIndep2 96.2 96.2
RegIndep3 93.5 94.2
RegIndep4 96.7 96.9
RegIndep5 95.8 95.8
(Data are classification rates, %. Data reduction: LLE. The alignment methods and features yielding CR>90% are highlighted in gray.)
Table 6. Semi-factorial analysis: pre-processing and dimensionality reduction methods.
(a) 104-fold cross-validation
Foot Pre-processing Dimensionality reduction
None 91.4 95.4
Decorr 92.3 96.2
RegMunCont 99.0 99.4
RegIndepMean 94.2 94.2
None 92.9 94.8
Decorr 90.2 93.9
RegMunCont 99.2 99.8
RegIndepMean 95.0 93.7
(b) Leave-one-out cross-validation
Foot Pre-processing Dimensionality reduction
None 91.4 95.4
Decorr 92.5 96.2
RegMunCont 99.0 99.4
RegIndepMean 94.2 94.4
None 93.1 95.0
Decorr 90.4 93.9
RegMunCont 99.2 99.8
RegIndepMean 94.2 94.4
(Data are classification rates, %. Combined pre-features: P100 and P80. The best-performing methods are highlighted.)
Figure 1. Description of plantar pressure data for a single step. (a) Pressure image time series; percentages
indicate normalised time (% stance). (b) Pixel time series; dark grey, black, and light grey trajectories indicate
pixels whose maxima were reached in the first, second, and final thirds of stance phase, respectively. (c) Pre-
features for an example pixel time series (see Table 2 for variable descriptions). (d) Pre-features, when computed
across all pixels.
Figure 2. Maximal pressures (P100) for the chronologically first twelve subjects; averaged across five trials.
Figure 3. Spatial alignment example, first subject. Top, middle, and bottom rows depict the original,
decorrelated, and registered images, respectively; here the registration template was RegMunCont (Table 2). The
thick dark outline depicts the cross-trial mean.
Figure 4. Classification rate (CR) for all 104 subjects using the P100 pre-feature. See Tables 3 and 4 for
alignment and dimensionality reduction and alignment method descriptions.
Figure 5. Classification rate (CR) as a function of embedding dimension (alignment: RegMunCont, pre-feature:
P100, reduction: LLE).
Figure 6. ‘Follow-up’ test results. (a-b) Number of correctly classified images (out of five) for the left and right
feet (light and dark bars, respectively) and for all ten follow-up subjects (s01…s10). Numbers in the white boxes
indicate the number of years between collection of the follow-up and main datasets. (c) Left foot P100 images
for subject 6, mean across five trials. The ‘Target’ image is from the main dataset.
Figure 1. Description of plantar pressure data for a single step. (a) Pressure image time series; percentages indicate
normalized time (% stance). (b) Pixel time series; dark grey, black, and light grey trajectories indicate pixels whose
maxima were reached in the first, second, and final thirds of stance phase, respectively. (c) Pre-features for an example
pixel time series (see Table 2 for variable descriptions). (d) Pre-features, when computed across all pixels.
Figure 2. Maximal pressures (P100) for the chronologically first twelve subjects; averaged across five trials.
Figure 3. Classification rate (CR) for all 104 subjects using the P100 pre-feature. See Tables 3 and 4 for alignment and
dimensionality reduction and alignment method descriptions.
Figure 4. Classification rate (CR) as a function of embedding dimension (alignment: RegMunCont, pre-feature: P100,
Figure 5. ‘Follow-up’ test results. (a-b) Number of correctly classified images (out of five) for the left and right feet (light Download full-text
and dark bars, respectively) and for all ten follow-up subjects (s01…s10). Numbers in the white boxes indicate the number
of years between collection of the follow-up and main datasets. (c) Left foot P100 images for subject 6, mean across five
trials. The ‘Target’ image is from the main dataset.