Information Technology Reference
In-Depth Information
(a)
(b)
Fig. 5. Volume whose slices are consecutive silhouettes of one person (a) or two persons
(b) crossing the curtain over time
Model-based approaches process sequences of images to estimate the param-
eters of an explicit gait model. These estimated values are then used to recover
the identity of the walking human. It should be noted that, in our case, these
silhouettes need to be reconstructed from the contour-related information given
by the two radial sensors.
These methods often need high definition images in order to work properly
which is a major drawback for our application since the laser sensors only provide
274 points per scan for an angular aperture of 96°. Furthermore, they exhibit
a significantly higher computational cost than silhouette-based techniques. This
is also a serious issue since real-time processing is required in our application.
Silhouette-based approaches do not rely on any explicit model for the walking
human(s). These techniques extract signatures directly from series of silhouettes.
A simple approach is described in [7] where the areas of raw (re-sized) silhouettes
are used as a gait signature. In [8], the gait template of a walking human is
computed by averaging the corresponding binary silhouettes. The classification
is then achieved using a nearest neighbor technique.
The contours of silhouettes have been used in [9] and by Soriano et al. in [10]
where signatures are derived from series of Freeman encoding of the re-sized
silhouette shape. An angular transform of the silhouette is proposed in [11] and
is said to be more robust than the raw contour descriptions.
The gait signature of [12] is based on horizontal and vertical projections of the
silhouettes. The authors of [13] consider time series of horizontal and vertical pro-
jections of silhouettes as frieze patterns. Using the framework of frieze patterns,
they estimate the viewing direction of the walking humans and align walking se-
quences from similar viewpoints both spatially and over time. Cross-correlation
and nearest neighbor classification is then used to perform the identification of
the walkers. To get an increased robustness to differences between the training
and test sets, [14] proposes a technique that relies on frieze patterns of frame
differences between a key silhouette and a series of successive silhouettes.
 
Search WWH ::




Custom Search