Deep Learning Approaches for Real-Time Object Detection in Smart City Applications: A Comparative Study

TEST!

Abstract

Smart city initiatives worldwide increasingly rely on computer vision and deep learning technologies for urban management, security, and traffic optimization. This study presents a comprehensive comparative analysis of state-of-the-art deep learnng architectures for real-time object detection in smart city environments. We evaluated five prominent models: YOLOv8, Faster R-CNN, SSD, EfficientDet, and DETR across three critical smart city applications: traffic monitoring, pedestrian detection, and vehicle classification. Our experiments utilized a custom dataset of 50,000 images collected from 15 locations across Tashkent, Uzbekistan, over six months. Results demonstrate that YOLOv8 achieves superior performance with 94.2% mAP@0.5 and 67 FPS on edge devices, while maintaining 89.3% accuracy in challenging weather conditions. We propose an ensemble approach combin ing YOLOv8 for real-time detection and EfficientDet for high-precision scenarios, achieving 96.1% mAP with acceptable latency for smart city deployment. This research contributes novel insights into practical implementation challenges, including computational constraints, varying lighting conditions, and occlusion handling specific to Central Asian urban environments.