Enhancement to Self-learning License Plate Matching Algorithm: Starting Association Matrix
The self-learning license plate matching algorithm, which was developed by Oliveira-Neto et al. (2013), has been tested and proven satisfactory at matching license plates between two LPR stations for traditional applications, but requires enhancements to be proficient enough for short-term data collection (equivalently, small sample size), which needs faster learning speed. The proposed improvements are to the initial process of the matching algorithm; more specifically, the starting association matrix. Until the algorithm has learned enough, the starting association matrix lays down all groundwork for determining the edit distances that are a crucial part for establishing whether two license plates are a match. In order to guarantee the fastest and most effective attainment of accurate edit distances, the better starting association matrix must be chosen. The enhancements are aimed at improving the true matching rate, false matching rate and learning speed. Twelve potential starting association matrices were evaluated. The results reveal that several starting association matrices helped the algorithm perform much better compared to an identity matrix and the selection of a starting association matrix is dependent on the application of the collected travel information. The top starting association matrices, after one learning iteration, achieved matching rates of 97% with 1.3% false matching for high learning speed (25 license plates).
