Carter, John
ON MEASURING TRAJECTORY-INVARIANT GAIT SIGNATURES
John N. Carter and Mark S. Nixon
Department of Electronics and Computer Science,
The University, Southampton, UK.
inc(@ecs.soton.ac.uk, msn@ecs.soton.ac.uk
KEY WORDS: Gait Recognition, Biometrics, Invariance, Computer Vision
ABSTRACT
Biometrics are increasingly important as a means of personal identification, and as such automatic gait analysis is
emerging as one of the most promising new techniques for non-contact subject recognition. There are many problems
associated with obtaining a gait signature automatically, in particular the effects of footwear, clothing and walking
speed. Furthermore, laboratory studies have constrained subjects to walk in a plane normal to the camera's view and
have ignored the effects of pose. Methodologies based on modelling human walking offer the opportunity to develop
analytic pose compensation techniques; here we develop a new geometric correction to the measurement of the hip
rotation angle, based on the known orientation to the camera, using the invariance properties of angles under geometric
projections. We present experimental results showing the application of our corrections to geometric targets and a real
human walker. We also indicate that it is possible to derive the corrections from the gait data itself. As such we
demonstrate that it is indeed possible, by geometric analysis, to provide invariant signatures for automatic gait
recognition.
1 INTRODUCTION
Personal identification is becoming an ever-important issue in everyday life. The need for personal security, access
control and identification is increasingly significant in individual and national political agendas. Further, the incidence
of fraud and impersonation is rife. For example, US and UK welfare fraud costs billions of pounds per year, whilst one
credit card company, MasterCard, estimates that it alone loses $450 million per year'. These issues can be addressed by
an effective person identification system. There are many strategies for personal identification, based on knowledge
(e.g. passwords), on possession (e.g. identity cards) or on some unique property of the individual, a personalised
measurement or biometric.
Any human physiological characteristic is potentially a biometric, provided it is universal, unique, permanent and
collectible’. That is, everybody has this property, it is unique to the individual, does not change over time and is
measurable. Biometrics range from established methods such as fingerprints, through voice and face recognition, to new
and emerging techniques such as iris identification.
No biometric is perfect, many suffering from social and practical problems, for example the need to make physical
contact when using fingerprint systems, or the potential social embarrassment when interrogating a public voice
recognition system. Unlike fingerprints and signatures, biometrics that need no subject contact (such as face
recognition) are more acceptable to users but can be limited by practical issues (such as face visibility). Gait recognition
is one of the newest of the emergent biometrics, and has the potential to overcome many problems. It is a non-contact
biometric, requiring no subject interaction. Also, in general the whole body presents a larger and more accessible target
than just the human face and in many applications scenarios, especially those involving serious crime, it is likely that
the face will be wholly or partially obscured whereas the gait will not. For these reasons gait now attracts research
interest. However, as yet there has been no study of the appropriate basis for measuring gait for purposes of automated
recognition: that is the subject of this paper.
We first discuss the bases for gait measurement. The two main approaches are 'statistical'
discuss how the model-based approaches have better generalisation capability (to practical ap
section. In Section 3 we then investigate the principal basis of the model on which gait meas
how the variation in trajectory effects perceived gait signature. These observations are
and 'model-based' and we
plication), in the following
urement is formed, showing
confirmed by experimental
114 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000.
studie
drawn
There
identi
exam]
attenti
Early
differ
others
biome
gener.
forens
subje
The t
appro
ofan
taken
Analy
succe:
gait re
statist
proce
The 2
rotati
motio
freque
motio
Fig
ang
the lii
path
requii
Cons
came
motic
grour
condi
nume
statis
that a