next up previous
Next: Visual Cues Up: Background and Theory Previous: Background and Theory

Voting

Voting methods in general, deal with $ n$ input data objects, $ c_i$, having associated votes/weights $ w_i$ ($ n$ input data-vote pairs $ (c_i, w_i)$) and producing the output data-vote pair $ (y, v)$ where $ y$ may be one of the $ c_i$'s or some mixed item. Hence, voting combines information from a number of sources and produces outputs which reflect the consensus of the information. The reliability of the results depends on the information carried by the inputs and, as we will see, their number. A cue is formalized as a mapping from an action space, $ \textbf{A}$, to the interval [0,1], $ c~:~\textbf{A}~\rightarrow~[0,1]$. This mapping assigns a vote or a preference to each action $ a
\in \textbf{A}$, which, in the context of tracking, may be considered as the position of the target. These votes are used by a voter or a fusion center, $ \delta (\textbf{A})$. Based on the ideas proposed in [7], [8], we define the following voting scheme:


- Weighted Plurality Approval Voting For a group of homogeneous cues, $ \textbf{C}=\{c_1, \dots, c_n\}$, where $ n$ is the number of cues and $ O_{c_i}$ is the output of a cue $ i$, a weighted plurality approval scheme is defined as:

$\displaystyle \delta(a) = \sum^{n}_{i=1} \hspace{1mm} w_i \hspace{1mm} O_{c_i}(a)\vspace{-4mm}$ (2)

where the most appropriate action is selected according to:

$\displaystyle a'=\operatorname{arg max}\{\delta(a)\vert a \in \textbf{A} \}$ (3)


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