On-Board License Plate Obfuscation System

The widespread deployment of networked surveillance cameras has increased concerns about the exposure of personally identifiable information, particularly vehicle license plates, before any privacy protections are applied. Many monitoring systems stream or store raw video and perform anonymization off-device. This creates windows during which sensitive imagery can be intercepted, misused, or lead to regulatory noncompliance. Despite these concerns, current solutions do not ensure that license plate data is anonymized directly at the point of capture, thereby minimizing the risk of malicious interference or data misuse. This project presents an FPGA-based edge-processing system that guarantees license plate anonymization at the point of capture, ensuring that unprotected imagery never leaves the device. Our design implements a complete edge-processing pipeline on the ZUBoard 1CG that captures video from a parallel-interface camera, identifies plate regions using a lightweight NanoDet-Plus or YOLOv5 neural network, and immediately applies configurable obfuscation. The anonymization module supports both irreversible black-box masking for default privacy protection and reversible key-based bit-shuffling for authorized recovery under controlled conditions. A streaming architecture allows the system to maintain real-time performance while integrating preprocessing, inference acceleration, post-processing, and output formatting on a single FPGA platform. The prototype targets ≥25 frames per second with <100 ms latency while preserving detection accuracy for campus and roadway environments. Experimental evaluation will measure detection precision, timing performance, and resistance to automated recovery. This project demonstrates that privacy-preserving video surveillance can be achieved at the edge without reliance on cloud processing while balancing public safety and individual privacy.