Real-Time UAV-Based License Plate Recognition via Analog Video Transmission and Closed-Loop Cloud–Ground Incremental Learning

Unmanned Aerial Vehicles (UAVs) are increasingly deployed for real-time traffic monitoring and surveillance; however, their onboard processors cannot support continuous high-accuracy visual inference. To address this limitation, we propose a real-time cloud–ground collaborative license plate recognition (LPR) system driven by analog video transmission and continuous cloud-based model improvement. A microcontroller-based quadcopter streams low-latency analog video to a ground station, where YOLOv8 performs real-time license plate detection and EasyOCR performs text extraction. Instead of relying solely on cloud inference, the system uploads low-confidence or misclassified samples to a cloud server, which performs periodic fine-tuning to adapt the model to analog-induced noise, varying illumination, and real-world flight conditions. Updated models are then redeployed to the ground station, forming a closed feedback loop for incremental learning. Comprehensive experiments quantify end-to-end latency across analog transmission, ground-side inference, and cloud retraining cycles, and show that the proposed architecture reduces the full-plate failure rate from 18.7% to 9.6% over three retraining cycles while holding measured end-to-end delay within 169–209ms. The system offers a scalable, energy-efficient, and self-improving solution for UAV-based visual surveillance and intelligent transportation applications