Digital Signal Processing Reference
In-Depth Information
The remainder of this paper is organized as follows. An overview of the vehicle
tracking algorithm based on block level motion vector is stated in Section 2. The pro-
posed vehicle tracking algorithm based on pixel level motion vector is described in
Section 3. The experiment results are shown in Section 4. The Conclusion is con-
tained in Section 5.
2
Block Level Motion Vector
An image of 640×480 pixels is divided into 80×60 blocks, where each block consists
of 8x8 pixels. The blocks at which the difference between intensities and those at the
background image is smaller than a threshold value is regarded as background blocks,
and the other blocks is regarded as foreground blocks. A vehicle is described by a
group of foreground blocks with the same vehicle ID. Blocks in each frame image are
given background ID at the beginning. Vehicles are shifted from the current image to
the next image based on motion vector estimated by block matching.
Fig. 1.
Different actual matching blocks enjoy the same approximate matching block
Due to dividing an image into non-overlapping blocks, it is necessary to normalize
motion vector form pixel level to block level, in other words, both the horizontal and
vertical components of a motion vector must be rounded to integral multiples of side
length of block. For convenience, the block which is shifted from last frame image
based on pixel level motion vector is called the actual matching block, and the block
which is shifted from last frame image based on block level motion vector is called
the approximate matching block. When different actual matching blocks enjoy the
same approximate matching block as shown in Fig.1, only one of them can be re-
served. Therefore, some vehicle blocks are lost in shifting processing. Blocks
reduction can cause vehicle location inaccuracy, and even missing vehicle halfway.
