next up previous
Next: Action Fusion Approach Up: Implementation Previous: Initialization


Response Fusion Approach

After the target is located, a template is initialized which is used by the correlation cue. In each frame, a color image of the scene is acquired. Inside the window of attention the response of each cue, denoted $ O_i$, is evaluated, see Fig. [*]. Here, x represents a position:
Color - During tracking, all pixels whose color falls in the pre-trained color cluster are given value between [0, 255]:

\begin{displaymath}\begin{split}&0 \leq O_{color}(\textbf{x},k) \leq 255 \hspace...
...bf{z}}_{k\vert k-1} + 0.5\textbf{x}_w] \end{split}\vspace{-2mm}\end{displaymath} (8)

where $ \textbf{x}_w$ is the size of the window of attention.
Motion- Using (Eq. [*]) and (Eq. [*]) with $ \Gamma=10$, image is segmented into regions of motion and inactivity:

\begin{displaymath}\begin{split}& 0 \leq O_{motion}(\textbf{x},k) \leq 255-\Gamm...
...mm} \hat{\textbf{z}}_{k\vert k-1}+ 0.5\textbf{x}_w] \end{split}\end{displaymath} (9)

Correlation - Here, the output is given by:

\begin{displaymath}\begin{split}& O_{SSD}(\textbf{x},k)=255e^{(-\frac{\Bar{\text...
...in [-0.5\textbf{x}_w, \hspace{1mm} 0.5\textbf{x}_w] \end{split}\end{displaymath} (10)

with Gaussian centered at the peak of the SSD surface.
Intensity variation - Here, (Eq. [*]) is used. If a low variation is expected, all pixels inside a $ m\times
m$ region are given values (255-$ \sigma$). If a large variation is expected, pixels are assigned $ \sigma$ value directly. $ m$ depends on the size of the window of attention with $ m=0.2\textbf{x}_w$:

\begin{displaymath}\begin{split}&0 \leq O_{var}(\textbf{x},k) \leq 255 \hspace{3...
...m} \hat{\textbf{z}}_{k\vert k-1} + 0.5\textbf{x}_w] \end{split}\end{displaymath} (11)


Fusion:
The responses are integrated using (Eq. [*]):

$\displaystyle \delta(\textbf{x}, k)=\sum_i^{n}w_i O_i(\textbf{x}, k)\vspace{-2mm}$ (12)

However, (Eq. [*]) can not be directly used since there might be several pixels with same number of votes. Therefore, this equation is slightly modified to accommodate for this:

$\displaystyle \delta^{'}(\textbf{x},k)=\left\{ \renewedcommand{arraycolsep}{2pt...
...0.5\textbf{x}_w]\} \end{array} \\  0 & : & \text{otherwise} \end{array} \right.$ (13)

Finally, the new measurement $ \textbf{z}_k$ is given by the mean value (first moment) of $ \delta^{'}(\textbf{x},k)$, i.e., $ \textbf{z}_k=\Bar{\delta}^{'}(\textbf{x},k)$.


next up previous
Next: Action Fusion Approach Up: Implementation Previous: Initialization
Danica Kragic 2002-12-06