Deep Learning Architectures for Real-Time Object Detection

Abstract

Real-time object detection—extracting category labels and bounding boxes at video frame rates—has become foundational for autonomous systems, industrial automation, smart cities, and augmented reality. This article surveys mainstream deep learning architectures with an emphasis on latency–accuracy trade-offs, deployment constraints on edge hardware, and training strategies that enable sub-30 ms inference. We compare one-stage (e.g., YOLO, SSD, EfficientDet) and two-stage (e.g., Faster R-CNN) families, as well as transformer-based detectors (e.g., DETR). We synthesize best practices for model compression (quantization, pruning), label assignment, loss design (IoU-aware), and data pipelines to maintain high mean average precision (mAP) while achieving high frames per second (FPS). We conclude with a practical checklist for selecting architectures under compute, energy, and memory budgets, and highlight open challenges such as long-tail detection, multi-sensor fusion, and trustworthy deployment.