19-11-2012, 11:26 AM
Moving Object Detection and Tracking for Video Surveillance Usinga Spatio-Temporal Graph
1Moving Object Detection and Tracking.pptx (Size: 2.34 MB / Downloads: 61)
OBJECT DETECTION AND TRACKING
Compressed Domain
Compressed Domain processing avoids the full decoding and reconstruction of the video, which provides a potential for real time processing of multiple video streams on one server.
Compressed domain information relies on the motion vectors (MVs) in the code stream.
Algorithm works by comparing the phases of the DCT(Discrete CosineTransform) coefficients in a series of images.
Block Diagram Compressed domain Motion Detection
Parsing is the process of extraction of the MVs from the coded video stream. The parsing process requires entropy decoding and reconstruction of MV’s.
To remedy the problem of missing MVs , the
extracted motion field is pre-processed prior to analysis.
The missing MVs in a current
frame are interpolated from MVs from the previous and
the next frame
The MV is interpolated as an average of the MV from the past frame and an MV from the next frame pointing.
The interpolated motion vector is incorporated into the motion
field, thus providing a complete set of MVs over the 4 × 4
blocks.
The determination of which MVs correspond to true motion is based on spatial and temporal clues.
A measure, of how likely a motion vector is to represent a real motion, is introduced and called the confidence. The confidence can be thought of as a probability that a given MV is related to
a real moving object and is a number in the range (0, 1).
For the temporal confidence process we first define a reference MV) (RMV) for each MV in the currently processed frame. The RMV is assigned the value of the motion vector of the 4×4 block containing the center of the area pointed to by the MV of the current block.