Emergency Vehicle Detection in Urban Traffic: A Deep Learning Perspective

Abstract

Emergency vehicle detection in urban traffic is a critical component of intelligent transportation systems, directly impacting response times and public safety. This paper proposes a robust deep learning-based approach for real-time emergency vehicle detection using the YOLOv5 object detection framework. Urban environments present numerous challenges for accurate detection, including high vehicle density, varying lighting conditions, frequent occlusions, and diverse vehicle appearances. To address these, we curated a comprehensive dataset comprising real-world traffic scenarios with labeled emergency and non-emergency vehicles across multiple urban settings. The YOLOv5 architecture was fine-tuned with custom augmentation strategies such as random scaling, contrast adjustment, and noise injection to enhance generalization. Additionally, we incorporated an attention mechanism within the detection pipeline to better focus on distinguishing visual features of emergency vehicles such as sirens, lights, and color patterns. The proposed model achieved an attenuation-adjusted mean Average Precision (mAP) of 0.96, demonstrating superior detection accuracy and robustness under challenging conditions. Extensive experiments validate the model's capability to perform real-time inference with low latency, making it suitable for deployment in edge-based traffic monitoring systems. This work significantly contributes to the advancement of smart city infrastructure by enabling automatic emergency vehicle recognition, aiding in dynamic traffic signal control and improved emergency response coordination.