ALPR: The Devil is in the Details...

November, 2004

This is the second of two installments on Automatic License Plate Recognition systems by Jim Kennedy. He has been involved in the video industry since the mid-1970s, with a particular interest in closed-circuit video as it applies to traffic observation, and, over the past eight years, with vehicle license plate capture technology. In this installment, he discusses the details of problems with mismatched license plate numbers.

What are the actual chances for a mismatch at the exit? Using history that shows the highest probability ALPR errors (O and 0 or Q, B and 8, 1 and I, notably), there are two likely opportunities for this plate to be misread at either entrance or exit. It could be read as A8C123 or ABCI23. Arguably, it could be read as A8CI23 as well. Remember, it is now a captive number, having been read on the way in. At the point of exit, therefore, it is highly unlikely that it can be A8C123, as this plate never entered the facility. This, of course, assumes an accurately read license plate upon entry.
If it is misread the same at both entry and exit, it does not matter as it will be a match. A good ALPR system will virtually eliminate the misreads just suggested, because it knows the "geographic syntax rules" of plates likely to frequent a particular parking location. What would be the result if there was the occasional misread?
Let's consider what the probable odds are of having the correctly read ABC123 in the same 3,000-space parking facility as the incorrectly read A8C123 (if such a plate even existed?). To actually calculate those odds, which will be very, very long (the word gazillion comes to mind), you would have to determine if the geographic syntax rules -- the arrangement of numbers and letters allowed by a particular state or states -- would even allow for it to happen. It is surprising to see how many states have a unique syntax for their standard-issue license plate and how many share their syntax with only one other state or just a few other states. This is probably a poor example, as there may not even be a state that would use A8C123 as a standard-issue passenger car license plate. (Note the picture nearby.)
To compensate for the occasional misread character, and since we already know the "captive" (entry) plate, all that needs to be done is to compare the plate read at exit with any other plate in "inventory" that has:
* The same exact string of
characters
* Five characters the same, in the
same order
* Four characters the same, in the same order
If the character(s) that is not the same on both plates meets criteria such as one is an "0" and the other is an "O" or "Q" (or B & 8, I & 1, etc.), then it becomes even easier to assume with a very high degree of confidence that they are the same vehicle.
The certainty of it being the same vehicle as that which entered even when only four characters are the same, and in the same order, is very high. Consider the odds of two states, each with the same syntax and each having the same exact license plate number as a current issue, and then calculate the odds for the likelihood of both vehicle owners being at the same airport (or any other place) at the same moment in history? In the same lot? The mind boggles.
Consider one more thing that is very important in understanding the probability of two similar plates being in the same parking facility at the same time and therefore being confused at exit: One of them has to be a cheater -- someone who swapped a ticket or some other scheme where the ticket/transaction number does not match the license plate. Quite simply, it really doesn't happen in statistical terms.
Some end-users look to including state identification as a further means to eliminate the potential for misread plates being confused. State identification is something that good suppliers of ALPR hardware and software can provide, but in light of the odds of duplicate plate reads, it may be the wrong solution. It is important to most facilities to allow the traffic entering and leaving the parking facility to do so as rapidly as possible. Generally, it is thought that from the time a vehicle pulls a ticket from the dispenser at the entrance, it should take somewhere on the order of less than one second to have the barrier gate start to rise.
Adding additional algorithms allowing for the decoding of plate information to determine the state of issue, while adding a small amount of additional time as well as cost, would not seem to create any additional comfort level for increasing the odds of good license plate matching at exit. Something else to consider in such cases is the frequent software upgrades for those state-identifiable plates in an effort to keep up with changes made to plate characteristics, which seem to be made on a regular basis.

Jim Kennedy is President of INEX Technologies. Contact him through the company's Web site (www.inextek.com).