next up previous
Next: Weighting Up: Background and Theory Previous: Voting


Visual Cues

The cues considered in the integration process are:
Correlation - With the standard sum of squared differences (SSD), the position of the target is found at the lowest dissimilarity score:

$\displaystyle SSD(u,v)=\sum_{n}\sum_{m}\left[I(u+m, v+n) - T(m,n)\right]^2$ (4)

where $ I(u,v)$ and $ T(u,v)$ represent the grey level values of the image and the template, respectively.
Color - Color is represented by $ r$ and $ g$ component in the Chromatic Color space [1].
Motion - Motion detection is based on computation of the temporal derivative using image differencing:

$\displaystyle M\left[(u,v),k\right]= {\mathcal{H}}\left[\vert I\left[(u,v),k\right] - I\left[(u,v),k-1\right]\vert - \Gamma \right]\vspace{-1mm}$ (5)

where $ \Gamma$ is a fixed threshold and $ \mathcal{H}$ is defined as:

$\displaystyle {\mathcal{H}}(x)=\left\{ \begin{array}{ccc} 0 & : & x \leq 0\\  x & : & x > 0\end{array} \right.\vspace{-2mm}$ (6)


Intensity Variation - In each frame, the following is estimated for all $ m\times
m$ (details about $ m$ are given in Section [*]) regions inside the tracked window:

$\displaystyle \sigma^2 = \frac{1}{m^2}\sum_u\sum_v \left [I(u,v)-\Bar{I}(u,v) \right]^2\vspace{-2mm}$ (7)

where $ \Bar{I}(u,v)$ is the mean intensity value estimated for the window. For example, for a mainly uniform region, low variation is expected during tracking. The level of texture is evaluated as proposed in [11].


next up previous
Next: Weighting Up: Background and Theory Previous: Voting
Danica Kragic 2002-12-06