Digital Signal Processing Reference
In-Depth Information
A Multi-view Fuzzy Matching Strategy
with Multi-planar Homography
Jie Shao
1
and Nan Dong
2
1
Shanghai University of Electric Power,
Department of Computer and Information Engineering, Shanghai 200090
2
Chinese Academy of Sciences,
Shanghai Advanced Research Institute, Shanghai 201203
Abstract.
Occlusions and incorrect detection make it very difficult to combine
information from all views correctly in multi-view surveillance. As a result, we
proposed a fuzzy matching strategy using a multi-planar homography con-
straint. Different from conventional methods which determine relationships of
blobs based on their locations on the ground plane corresponding to the feet of
the people, our method employs a statistical strategy. First, we divide each tar-
get into several parts, and project them onto different planes in the space. Then
overlapped parts in different planes will be recorded. The optimal pairs appear
based on a voting strategy. Experimental results are shown in scenes from dif-
ferent view points and light conditions. The algorithm is able to accurately
match target blobs in all views. It is ideally suitable for conditions with not
enough features.
Keywords:
multi-view, fuzzy strategy, multi-planar homography.
1
Introduction
Multi-view collaborative sensing is one of the most challenging tasks in intelligent
video sensing. It is implemented based on single view target detection and multi-view
target matching. Multi-view target matching is one of the research directions of digital
image registration, and also the pre-processing stage of digital image mosaic and 3D
reconstruction. In the area of multiple targets sensing, a detected blob is no longer
guaranteed to be a single person, but may belong to several people. Even worse, a
person might be completely occluded by other people. As a result, with the help of
target matching, we could mix data from different view points together, to enrich our
information, and get more accurate results. As the example of a multiple pedestrian
tracking system mentioned in [1], a multiple blob tracker is employed and the position
of head is used to locate each pedestrian. Its performance decreases when targets are
close to each other and occluded. However, if targets are tracked in cameras of differ-
ent positions contemporarily, detection and tracking accuracy will be efficiently
increased [2].
Different from static image registration methods, the results of multi-view target
motion matching do not only rely on static feature matching, but their spatio-temporal
