Hardware Reference
In-Depth Information
Search the similar texture
Evaluate the ID distribution
41
41
41
41
41
Previous
Current
Previous
Current
y
x
y
x
17
17
time
time
17
17
41
41
41
41
17
17
17
41
17
41
41
41
17
17
41
41
17
41
41
Motion Vector MAP
Object MAP
Fig. 6.14
Overview of S-T MRF algorithm
An overview of the S-T MRF algorithm is shown in Fig.
6.14
. This algorithm
presumes the boundary of each object based on the motion vectors. First, the motion
vectors of each object are extracted by comparing the previous and current image
frames. The extracted motion vectors are mapped onto the motion vector maps.
With these motion vector maps, the object maps including the boundary informa-
tion of objects are also generated. In the process of generating these object maps,
the boundaries of objects are evaluated by a high-level algorithm and then opti-
mized. In this way, robust and stable object tracking is achieved by updating each
map in every image frame.
As shown in Fig.
6.15
, the S-T MRF can be divided into two layers, that is, object
map creation and motion vector extraction. In applications that use the S-T MRF
algorithm, the event detection algorithm is added in the application layer. The vol-
ume of data is reduced as the software level is elevated because the higher-level
layer does not need to handle the pixel data. These operations on each layer are
executed independently; therefore, the thread parallel processing can be applied.
Figure
6.16
is an overview of the thread parallel processing of the S-T MRF
application on the proposed SoC architecture. The motion vector extraction that
processes large-volume data is assigned to the MX core. The object map creation is
assigned to CPU#0, and event detection is assigned to CPU#1. Each thread com-
municates with the other threads by giving and receiving the information. For exam-
ple, the object map creation gives the pixel pointer to the motion vector extraction
and receives the motion vector map. Thus, effective thread parallel processing can
be achieved with this scheme.
Search WWH ::
Custom Search