Efficient and Accurate Vehicle Detection for Smart Cities Using Deep Learning

Abstract

Growing urban populations worldwide have intensified the need for automated, intelligent traffic monitoring solutions capable of operating under demanding real-world conditions. This work proposes a six-class vehicle detection system built upon the YOLO11s deep learning architecture, trained exclusively on Indian road imagery sourced from the Indian Vehicle Dataset (5,000 annotated samples spanning Car, Bus, Truck, Motorcycle, Bicycle, and Auto categories). Training was conducted over 30 epochs on an NVIDIA Tesla T4 accelerator, leveraging AdamW optimisation and Automatic Mixed Precision to maximise convergence speed and hardware efficiency. On the held-out test partition, the system recorded mAP@50 of 0.9936, mAP@50–95 of 0.9650, precision of 0.9980, and recall of 0.9928. End-to-end processing latency measured 8.8 ms per frame, confirming real-time deployment viability for smart city traffic management platforms.