Game Development Reference
In-Depth Information
Figure 5. Cumulative motion of body parts for different activity patterns:
Waving one hand, opening and closing arms, left-right movement.
motion of the body parts. We observe that different activity patterns can have
overlapping periods (same or similar patterns for a period) for some body parts.
Hence, the detection of start and end times of activities is crucial. To detect the
start and end time of a gesture, we use the gap between different gestures/
activities.
Eighty-six percent (86%) of the body parts in the processed frames and 90% of
the activities are correctly classified, with the rest considered the miss and false
classification. Details of the gesture recognition algorithm can be found in Ozer
& Wolf (2002a).
From 2D to 3D
In this subsection, we present our algorithm that generates a real-time 3D model
of the human body by combining 2D information from multiple cameras located
at 90 degrees to each other. We propose a new 3D method for activity
recognition in real-time. The proposed method that combines valuable 2D
information is fast, robust and accurate. It doesn't require any disparity map and
wire-frame information for model generation. We generate a global 3D ellipsoid
model of the human body parts from 2D ellipses and use the resulting 3D
information to verify the fit of the real body parts with the actual model. We can
process approximately 25 frames per second on each TriMedia board. Figure 7
shows the architecture for 3D model generation. Camera calibration and data
synchronization are main issues in data fusion from multiple cameras. Visual
reconstruction for virtual reality requires high accuracy, while real-time activity
recognition and trajectory estimation require high-speed techniques (Dockstader
& Tekalp's, 2001; Focken & Stiefelhagens, 2002; Schardt & Yuan, 2002). Note
that our system uses static cameras that do not require dynamic calibration.
Search WWH ::
Custom Search