I'm sorry, but I'm a bit confused by the example. Do you mean that all of this:

       my $foo = "$bar{foo}";
       $foo .= $baz->method(@foo);
       $foo .= join(',', @$bar{'foo','bar'});
[download]

represents a single string that has been assigned to some scalar variable -- e.g. it's bracketed by something like:

my $code_string = <<EOS;

... # your example here

EOS
[download]

In this case, it wouldn't be all that hard to cook up a suitable tokenizer-plus-token-classifier that breaks up the contents of $code_string and tucks the relevant pieces into an array, based on looking at the sigils and brackets that bound each word. It might take a few hours of careful study and testing to get it mostly right, and there may even be some foibles of perl syntax that could defy the best attempts (abundant obfus to beat your head against...), but that's no reason not to try for something that handles the vast majority of stuff. (Anyway, sorry but I don't know of anything off-hand that already does this.)

If, on the other hand, you mean to pull those various array elements out of the value being assigned to "$foo" in your example, then I just don't get it.

update: The key point to simplify the task would be deciding that you only care about variables whose names consist of  /(\S*)([\%\@\*\$])(\w+)(\S*)/ -- if that's your intention.

another update: That regex for spotting variable names would probably be more reliable after using B::Deparse on the target code, so that whitespace is normalized to a consistent style before you start.