Quantized YOLOv11 for Real-Time Road Object Detection in Smart City Environments

With the surge in urban vehicle density, smart cities are increasingly relying on AI-powered traffic management systems to address congestion challenges. In this work, we present a real-time road object detection framework based on the YOLOv11 model, tailored for smart city environments. Leveraging the BDDI00K dataset, we train and evaluate four YOLOv11 variants-Nano, Small, Medium, and Large-and analyze their trade-offs in terms of accuracy, inference speed, and model size. While larger models achieve higher detection accuracy (mAP@50 ranging from 49.6% to 62.62%), they incur higher latency and reduced frame rates. To facilitate deployment on edge devices with limited computational resources, we explore both dynamic and static post-training quantization techniques. Experimental results demonstrate that dynamic quantization maintains near-original accuracy with a minimal 0.6% mAP@50 loss while significantly improving inference time. Our findings highlight the potential of quantized YOLOv11 models for efficient and scalable traffic perception in real-time urban applications.