Gait Analysis and Human Motion Tracking - Intelligent Video Event Analysis and Understanding - page 41

we present results that show the improved efficiency in comparison with standard
methods. Finally, in Section 6 we summarize our findings.

2 An Overview of the Proposed Approach

The proposed approach has two phases, initialisation and continuous tracking .As-
sume that m frames are included in the initialisation phase and n frames ( n m )

in the whole sequence. The intrinsic camera parameters are known from previous
calibration. The purpose of the first phase is to acquire the long-term gait model of
the pedestrian.

>

Phase 1: Initialisation
Of the well-tested feature tracking algorithms, reviewed by Lepetit and Fua [17],
the Shi-Tomase-Kanade (STK) tracker [27] is used because it is still one of the most
accurate and reliable. The STK tracker has two stages, feature selection and fea-
ture tracking. The feature selection process computes the eigenvalues of a gradient
function for each pixel, comparing the result with a fixed threshold. In the published
algorithm, image features with higher eigenvalues were considered as good features
to track. However, we use the SUSAN (Smallest Univalue Segment Assimilating
Nucleus) operator instead to select an initial feature set in the first frame, due to its
known immunity to noise [28]. This is not used subsequently, unless the number
of tracked features falls below a pre-defined threshold in which case we are able to
re-initialise with new features.

Select C corner features in frame 1
For frames i=1:1:m-1 (STK-GLS)
Match features across frames i and
(i+1);
Estimate fundamental matrix;
Refine list of matched features;
Recover camera transformation and
scene geometry;
EndFor
Fit periodic gait model to camera trans-
formation data from frames 1 to m

As we track these features in successive frames, we minimise the residue error us-
ing an affine transformation, as in the STK algorithm, assuming the displacement is
small. However, we use a generalised least squares (GLS) algorithm [42] to recover
the frame-to-frame camera transformation and scene geometry as this is a robust
estimation technique that can derive good results when the error distribution is not
normally distributed [40]. Typically, m is chosen large enough to recover about two
complete strides, say m=50 frames for a 25Hz sampling rate. The number of fea-
tures, C , is a user parameter, typically set to 150. This leads to the recovery of a
temporary motion model with six degrees of freedom, i.e. the 3 displacements and