End-to-End Motorcycle Violation Detection with Region-Specific Automatic License Plate Recognition

Abstract

We present a region-aware, end-to-end motorcycle violation detection pipeline tailored to traffic conditions in Punjab, Pakistan, which integrates three YOLOv11-based components into a unified framework: motorcycle violation detection (MCVD) for helmet compliance and multi-rider analysis, license plate detection (LPD), and license plate character detection (LPCD). The system integrates lightweight object detection, BoT-SORT-based tracking, and character-level recognition, supported by a synthetic-toreal adaptation strategy that combines large-scale synthetic data with limited real samples. Two specific datasets are published, a 40,000-sample synthetic Punjab license plate dataset (PS-LPCD) and a 650-sample real-world dataset (PR-LPCD), which are publicly released in order to encourage research development and adaptation to the region. Class consolidation enhanced MCVD performance (weighted average F1 score: 0.77) and the LPD model performed at mAP₅₀ = 0.99. Two-stage fine-tuning on synthetic and real samples allowed LPCD to reach a character accuracy of ≈ 98% and a full-plate recognition rate of ≈ 90.7%, both surpassing EasyOCR and PaddleOCR, while also achieving lower per-plate latency. With a single motorcycle per frame, the sequential pipeline maintains a throughput of ≈ 9.5 FPS; the throughput reduces in scenes where there are many motorcycles. These findings indicate that synthetic pretraining, together with a small real fine-tuning, can be used to obtain a powerful, scalable, and region aware automatic license plate recognition (ALPR) system, which provides a reproducible method for detecting traffic violations across a variety of license-plate formats.