Game Development Reference
In-Depth Information
extraction of correspondences between images from different views and
projections of a 3D model while the later yields extraction of correspondences
between 3D articulated models and reconstructed visual input. Gavrila (1999),
Aggarwal & Cai (1999), and Moeslund & Granum (2001) presented overviews
of various methods used for articulated and elastic non-rigid motion detection,
human motion estimation, tracking, recognition, pose estimation and various
other issues based on human detection and activity recognition. More back-
ground information on gesture recognition can be found in Wu & Huang (1999),
Kohler & Schroter (1998) and LaViola (1999).
Luck et al. (2002) and Cheung et al. (2000) obtain 3D models of the moving
human body by extracting the silhouettes from multiple cameras. Although our
approach is similar from this point, the main algorithm used for the modeling is
different.
In Luck et al. (2002), the authors use a physics-based approach for tracking 3D
human models. The voxels obtained from the silhouettes exert attractive forces
on a kinematic model of the human body to align the model with the voxels.
Although this method enables very accurate modeling, it requires the human body
model to be acquired from a specific initialization pose.
Our main aim is to use 3D info for our HMM-based activity recognition in real-
time for different applications without requiring any specific pose or user
interaction. Another major difference is architectural, as the authors use one PC
where all the processing is done in a centralized way, while our architecture is
distributed with local processors.
In Cheung et al. (2000), the authors use a similar approach to perform 3D voxel-
based reconstruction by using silhouette images from multiple cameras. The
local processing is used only for silhouette extraction. The five silhouette images
are then sent to a host computer to perform 3D voxel-based reconstruction. The
proposed algorithm first reconstructs 3D voxel data and then finds ellipsoids that
model the human body. Our algorithm, on the other hand, first finds 2D ellipses
that model the human body via graph matching at each local processor and then
reconstructs 3D ellipsoids at a host computer. Note that 2D processing such as
pose estimation is independent of the 3D modeling and activity recognition.
In Cohen & Lee (2002), the authors propose an approach for capturing 3D body
motion and inferring human body posture from detected silhouettes. 3D body
reconstruction is based on the integration of two or more silhouettes and the
representation of body parts using generalized cylinders and a particle filter
technique. Each silhouette is also used to identify human body postures by using
support vector machine. In Kakadiaris & Metaxas (1998), a human body part
identification strategy that recovers all the body parts of a moving human is
employed by using the spatio temporal analysis of its deforming silhouette. 2D
shape estimation is achieved by employing the supervisory control theory of
Search WWH ::
Custom Search