Expanding license plate matching capabilities with secondary self-learning algorithm

To perform the postprocessing matching of license plates between two license plate recognition (LPR) stations, a self-learning matching algorithm was employed. The key component of this algorithm is an association matrix that is a unique translator, associating two LPR units in relation to how they may see or recognize the same characters differently, for a host of reasons. This association matrix consists primarily of high-confidence matches between two LPR stations estimated directly from a set of matched character pairs. The matching algorithm's performance decreases as the distance between the two LPR stations increases because of vehicles no longer traveling within an average travel time window, a low sample of vehicles traveling between the two LPR stations, or both. This paper proposes using a third LPR station to generate additional information to derive a better association matrix for an existing pair of LPR stations and thus replaces the existing learned-association matrix. In other words, the added LPR unit facilitates secondary or transferred learning to improve the matching performance of the first two units, even after the third LPR unit is subsequently removed. To evaluate this derived association matrix, the authors employed two simulations. They were to determine when the newly derived matrix should be used and to evaluate the overall performance of license plate matching.