is 
ms 
ng 
nd 
op 
ip 
ric 
al 
we 
ait 
ce 
ne 
Dy 
SS 
^» 
V 
id 
al 
ce 
ce 
ct 
et 
at 
2 
B 
al 
Carter, John 
  
studies, prior to observations on appropriate invariance properties, in Section 4, then further work and the conclusions 
drawn from this study. 
2 GAIT AS A BIOMETRIC 
There is a rich literature, including medical and psychological studies, indicating the potential for gait for person 
identification". Since people need to walk their gait is generally apparent. Additionally gait is hard to disguise. For 
example in a robbery, or other criminal activity, the need is to walk normally and unobtrusively rather than to attract 
attention. 
Early medical studies suggest that if all gait movements are considered, gait is unique’. In all it appears that there are 24 
different components to human gait, some are more variable than others and some are more difficult to measure than 
others, particularly out of the laboratory. Murray's work’ indicates that gait has the richness necessary for a successful 
biometric which, with its no contact nature and the high visibility of body parts, makes it a fruitful candidate for a 
general-purpose remotely-sensed biometric. Some potential applications for a gait based recognition system include 
forensics, to identify individuals involved in serious crimes, and security to analyse gait patterns to monitor unusual 
subject behaviour. 
The two main themes in current approaches to automatic gait recognition are statistical and model based. The statistical 
approaches derive a unique signature by computing a spatiotemporal pattern based on a sequence of segmented images 
of a moving person. Typically the shape of the body is reduced to a binary silhouette and some statistical measures are 
taken from the sequence of silhouettes. Techniques such a Principle Components Analysis and Linear Discriminant 
Analysis have been used to provide a statistical description of the sequence". These techniques have been very 
successful, achieving 10096 recognition rates, though on small subject populations. Most extant approaches to automatic 
gait recognition are statistical in nature, describing movement by optical flow or spatiotemporally”. However, as with all 
statistical measures, it is not clear exactly which features of gait contribute to the recognition and discrimination 
processes. 
The alternative approach is to base recognition on a physical model of human motion. Following Murray, the hip 
rotation angle has been modelled as a simple pendulum, whose motion is approximately described by simple harmonic 
motion®’. This assumes that the motion is basically sinusoidal in nature, repeating periodically with every step, with 
frequency, phase and amplitude closely related to the mechanics of the walking process. In fact, simple harmonic 
motion is insufficient to describe human motion, rather the motion is expressed as a Fourier series’. Gait recognition 
using this model-based approach relies on accurate feature 
identification, automatically or via human intervention and 
labelling. A straight line parallel to the upper leg is derived from 
each picture in a video sequence, and used to compute the hip 
rotation angle. These angles are then combined to produce a gait 
signature. Medical studies indicated that the significant 
information is contained in the low orders of the Fourier 
Sequence, and this has been borne out by achievement of 100% 
recognition rates with only the first two harmonics’. Again this 
was with a small number of subjects. 
Current laboratory based experiments indicate that gait is highly 
promising as a biometric. However, before it can be of practical 
use the many different effects that may perturb and influence 
recognition rates must be quantified. In particular, the view 
angle and camera positions are all well controlled in laboratory 
experiments, usually by forcing the subject to walk normal to 
the line of sight of the camera. This will not be true in almost all-real world applications, where the angle of a subject's 
path with respect to the camera will be totally uncontrolled. This mandates that some form of invariance or correction is 
required to normalise the signatures of walking subjects to be independent of pose. 
    
Figure 1 Subject walking, left to right, at an 
angle of 20° to the camera. 
Consider a fixed video system, monitoring a person walking at a fixed angle to the view direction of the camera. If the 
camera optics have been fully calibrated, and the scene geometry is known, then it is entirely possible to reconstruct the 
motion of the walker. In principle this can be expedited by assuming that the person is walking perpendicular to the flat 
ground plane, and rotating the co-ordinate system such that labelled features appear as if they were viewed in laboratory 
conditions. This is not difficult to achieve but does impose severe constraints on generalisation capability, and the 
numerical processing required may introduce systematic errors to the derived gait signature. Clearly, this can affect 
statistically based approaches more than model-based ones, especially if it is possible to develop a system or algorithm 
that allows simple corrections to be made to the model (the hip rotation angle). 
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B5. Amsterdam 2000. 115