Of the people who say that they are undecided in a poll, how many are really undecided, and how many will claim to be undecided while secretly leaning one way or another? Even the respondents might not know for sure, but some Canadian psychology researchers have applied a fascinating way to find out (h/t to the Frontal Cortex):

In the study, researchers asked 132 residents of Vicenza, Italy, where they stood on a controversial ballot issue being debated at the end of last year: the expansion of a local United States military base. The voters also performed a computer test, in which they had to quickly categorize words flashing by on the screen as positive or negative; the words were sometimes paired with photos of the military base. By tracking the number of errors and the time it took to answer, the test estimated a person's implicit positive or negative associations with the base.

The researchers then had the voters come in a week later and found that of the 33 who had declared themselves undecided, 9 now were in favor and 10 were opposed. And their scores on the computer test the previous week predicted which way they would turn, said one of the authors, Bertram Gawronski of the University of Western Ontario.

When we measure the number of undecideds, we have to rely on self-reporting, which can be inaccurate--either people won't know their intentions (despite having an opinion on the question at hand) or else they will simply not want to report them.  For example, some people have a slight taboo about saying whom they're planning to vote for, even in an anonymous survey; they will therefore say undecided, even if they have an opinion.  Other people might simply want the social prestige that comes from being "open-minded" and "not voting the party line".  Whatever the motivation, it's a real problem, which is why some pollsters give a push to those who claim indecision ("Well, if the election were held today, whom would you lean towards?"). 

Due to expense and difficulty in setup, I doubt that the technique described in the article will ever come into widespread use amongst political pollsters.  But it's worth looking into as a technique in focus groups, although there would be sample size issues.  And aside from that, it's simply a neat way of solving a perennial problem in survey research.