Deep learning–based real time speed limit sign detection with YOLOv12 on edge AI platforms for embedded ADAS

Abstract

This research examined a real-time speed-limit sign detection framework based upon deep learning using the YOLOv12 neural network, optimized for the use of small edge devices that are embedded advanced driver assistance systems (ADAS). You only look once version 12 (YOLOv12) achieved a remarkable detection performance, while maintaining efficient computation, utilizing significantly optimized lightweight attention modules with an R-ELAN backbone capable of small and partially occluded detection. A custom dataset comprising 23,000 annotated images was prepared and augmented to ensure robustness under varying conditions. Model training utilized quantization-aware techniques and optimization via TensorRT and ONNX Runtime. Deployment and performance were rigorously evaluated on resource-constrained edge platforms, specifically NVIDIA Jetson Nano and Raspberry Pi 5. Experimental results demonstrated exceptional detection performance, achieving a precision of 99.0%, recall of 99.1%, and mean average precision (mAP@50) of 99.2%, confirming YOLOv12’s suitability for reliable, real-time ADAS implementation in intelligent transportation and autonomous vehicles.