Digital Signal Processing Reference
In-Depth Information
motion features in most of the cases. For example, in [3], they applied planar homo-
graphy to project blobs of each view images onto the ground plane, and then particle
filtering works in the 3D space. This kind of conventional methods is based on the
assumption that feet of all the pedestrians are on the ground all the time. It is not real
in some situations, like the moments of jumping, running and so on. Besides, as the
area of feet is very small in most of the images, a foreground blob may not have accu-
rate feet area after target detection, or a blob could be bigger or smaller than its real
size. Arsic D.[4] proposed an idea of m
ulti-Layer homography
for a more precise 3D
reconstruction of the scenery. Khan and Shah [5,6] proposed their multi-view image
association method from a novel point. One of the views is set as the reference view,
and then the foreground likelihood maps from all the other views are warped on to the
reference view to produce a 'synergy map', which is used to obtain pixels that
represent ground plane locations of people in the scene. It effectively decreases com-
putational costs in 3D tracking but does not increase matching accuracy. In addition
to these achievements, Javed et al. [2] obtain relationships between objects based on
FOV of different cameras.
In most existing cases, methods based on planar homography are not reliable, so
they are often followed by modification approaches based on color distribution or key
points matching. Cai et al. [7] selected N points belonging to the medial axis of the
upper body as the feature for tracking, and then tracked based on geometric parame-
ters like position and velocity. Delamarre et al. [8] used only contours of the sil-
houettes of the person to form its 3D geometric model. With the help of the extension
of Annealed Particle Filtering, Deutscher et al. [9] tracked complex human move-
ments without the use of extra constraints such as labeled markers, pose assumptions,
restricted movement or color coding. Besides, in [10], kernel color histogram
based mean shift method is used in multiple camera tracking. But color histogram is
unstable in the condition of light variation, and it is also not suitable for blobs with
small size.
Based on the above considerations, if we make full use of spatio-temporal informa-
tion composed of all views, not only modification process like feature matching could
be largely simplified or even be omitted, but the performance of the whole matching
method could also be improved. Borrowed the idea of multi-planar homography from
[4], and associated it with multi-part color histogram confidence mentioned in [11],
we proposed a multi-part fuzzy matching method based on multi-planar homography
in this paper. It is cooperated with a voting strategy to obtain the final result. The
method is composed of three steps: 1. foreground blob modeling; 2. multi-part fuzzy
projection; 3. voting strategy.
2
Foreground Blob Modeling
After our pixel-based background subtraction, blob analysis is performed on the
background-subtracted images based on morphological operators to connect clustered
pixels and remove isolated noise like ``salt \& pepper'' noise and diminutive clusters
of pixels. The details of the process are presented in [12].
