Automatic License Plate Detection and Recognition in Peru with AI: Training and Evaluation with COCO Metrics for Technical Guidelines in Traffic Management and Public Safety

We present a training-and-evaluation study of an ALPR pipeline (plate detection, LPD, plus plate reading, LPR) for the Peruvian domain, with emphasis on reproducibility and usefulness for technical guidelines in traffic management and public safety. Detection was trained with YOLOv8 (50 epochs, 640 px) under a COCO-style protocol; we assessed five training sizes N=500, 1000, 1500, 2000, 2514 while keeping validation/test fixed. The best epochs reached mAP@[0.50:0.95] = 0.4717, 0.6197, 0.6894, 0.7237, 0.7305, respectively; for N=2514 we obtained [email protected] = 0.9563, Precision = 0.8405, Recall = 0.9384, showing steady improvements with diminishing returns beyond ~1,500–2,000 samples. The operating point was chosen from the F1–confidence curve at conf = 0.721, consistent with PR curves. For LPR, we propose evaluating character accuracy (1–CER), Levenshtein distance, and full-plate match, in addition to end-to-end accuracy (LPD ∧ LPR). Results are discussed against reference benchmarks and accompanied by export/inference practices to facilitate institutional transfer. This work provides a reproducible basis for selecting sample sizes, thresholds, and performance expectations for ALPR in Peru.