A Virtual Curtain for the Detection of Humans and Access Control - Advanced Concepts for Intelligent Vision Systems - page 104

In our application, we apply an approach similar to silhouette-based gait
recognition. First we extract surfacic features from a single silhouette. Then
we aggregate features over time to obtain a temporal signature that is used to
identify the ongoing crossing scenario.

3.1 Feature Extracted from the Intersection between an Object and
the Virtual Curtain

To characterize reconstructed silhouettes, we use the notion of cover by rectan-
gles . The cover by rectangles is a morphological descriptor defined as the union
of all the largest rectangles that can fit inside of a silhouette. The whole idea is
described in [15].
From the cover of a silhouette, many features can be extracted to build a sil-
houette signature. Features that could be considered to characterize the dataset
are:

- The set of the enclosed rectangles (that is, the cover itself).
- The maximum width (or height) of the rectangles included in the cover.
- Histogram of the widths (or heights) of the rectangles included in the cover.
- 2D histogram of the widths and heights of the rectangles included in the
cover.
- The horizontal or vertical profile of the silhouette.

Due to the unusual shape of the silhouette, there is no prior art about the best
suited characteristics. Therefore, we fall back to proved intra-frame signatures
that were considered in [16] for gait recognition. They are:

- The 2D histogram of the widths and heights of the rectangles included in
the cover (denoted as

G W × H ( i,j )), and
- The concatenation of the histogram of the widths and the histogram of the
heights of the rectangles included in the cover (denoted as G W + H ( i,j )).

Note that in order to build histograms, we partition the widths and heights of
the rectangles respectively into M bins and N bins. The best values for M,N
are discussed later.

Temporal features. The full signature is constructed as a combination of
intra-frame silhouette signatures. Its purpose is to capture the time dynamics
of the moving object crossing the door. In our application, the time dynamics
may be very important. One of the proposed solution to handle the temporal
evolution of a shape is to normalize the gait cycle, like in [17].
Like for gait recognition, this poses a problem in that the classifier will delay
its answer until the end of the sequence. An alternative solution is to normalize
the sequence by parts. Another approach to consider is to learn several speeds
during the database set-up, and use the global normalization as a fallback or
confirmation step.
Our approach is much simpler and provides results similar to results obtained
with other approaches. Our inter-frame spatio-temporal signature (denoted as