Multiview Image Segmentation and Video Tracking - Video Segmentation and Its Applications - page 135

Fig. 5.12 Detection and tracking using single-layer homography: projection from multiple views
to the ground plane of top-down view

views are homographically mapped into the top view ground plane and the inter-
sections are estimated as the feet point. Multi-hypothesis trackers are established
using particle filters for robust tracking. A similar work is reported by Khan and
Shan [ 17 ], where tracking multiple people is solved by graph cut segmentation of
spatiotemporal feet blob volume.

5.3.2.2 Multi-Layer Homography

To enhance the accuracy of localization and reduce the false alarm, it can exploit
multiplanar homography constraint to combine projections in multiple layers that
are parallel to the ground plane (Fig. 5.13 ).
Tong et al. [ 41 ] proposed a multicamera approach for multipeople localization
using multiplanar homography constraint. Foreground regions are segmented us-
ing Gaussian mixture model-based background suppression for each view and each
frame, which are warped to the reference view to get target section on the plane.
Five-planar homography from ground plane to head plane is adopted to gather all
plane information to final overlooking view, where the people are clustered for
localization. A homography framework in [ 1 ] is developed for multicamera track-
ing. Foreground blobs are extracted using graph-cut based foreground sub-traction.
Applying multi-layer homography results in the increasing reliability of localization
by transforming the foreground map of reference view with multiple layers, where
the feet positions are detected to indicate the coherent foreground regions. Instead of
tracking footage of people, Eshel and Moses [ 9 ] worked on tracking people's head