Real-Time Ambulance Detection on Resource-Constrained Edge Devices Using YOLOv5 and ONNX Runtime

Abstract

Rapid response of Emergency Medical Services (EMS) is critically affected by urban traffic congestion, where conventional siren-based and manual traffic control mechanisms often fail to ensure timely right-of-way for ambulances. To address this limitation, this paper presents a real-time, vision-based ambulance detection system optimized for deployment on resource-constrained edge devices. The proposed system employs a fine-tuned YOLOv5s object detection model trained on a custom ambulance dataset and optimized using the Open Neural Network Exchange (ONNX) framework for efficient CPU-based inference. The optimized model is deployed on a Raspberry Pi (64-bit) platform using ONNX Runtime and integrated with a live IP camera stream for continuous detection. Experimental results demonstrate a Mean Average Precision (mAP@0.5) of 91.3% and a real-time inference speed of 2–5 FPS on the edge device. A comparative evaluation shows that ONNX Runtime significantly outperforms native PyTorch inference on the same hardware. The results demonstrate the practical feasibility of deploying a vision-based ambulance detection system on CPU-only edge devices without hardware accelerators, making the solution suitable for cost-sensitive Intelligent Transportation System deployments. The work primarily validates edge-level detection feasibility, with traffic control components evaluated via simulation. This work is positioned as an applied feasibility and deployment-oriented study rather than a novel algorithmic contribution. The study primarily evaluates deployment feasibility and runtime performance on CPU-only edge hardware.

Introduction

The text presents a real-time, edge-based ambulance detection system aimed at improving Emergency Medical Services (EMS) response in urban areas, where traffic congestion often delays ambulance arrival during the critical “golden hour.” Traditional methods—sirens, GPS, V2X communication—face limitations due to noise, infrastructure dependence, or hardware costs. Vision-based detection using deep learning, particularly YOLOv5, offers a robust, infrastructure-independent solution suitable for real-time applications.

The proposed system deploys YOLOv5s on a CPU-only Raspberry Pi, optimized via ONNX Runtime, providing a cost-effective solution without relying on GPUs or accelerators. Key contributions include:

1. Implementation and evaluation of a YOLOv5s-based ambulance detection system on a Raspberry Pi.

2. Comparison of PyTorch CPU inference versus ONNX-optimized inference.

3. Demonstration of real-time feasibility on low-cost edge devices.

System Design & Methodology:

• Dataset: ~1,200 images of ambulances under varied traffic and environmental conditions, annotated in YOLO format, split 70% training, 20% validation, 10% testing. Data augmentation improves generalization.

• Model Training: YOLOv5s with COCO-pretrained weights, SGD optimizer, 150 epochs, 640×640 input resolution.

• Model Optimization: Conversion to ONNX format for faster CPU inference using ONNX Runtime.

• Edge Deployment: Raspberry Pi processes live video frames, detects ambulances, and generates alerts. Post-processing includes confidence thresholding and Non-Maximum Suppression.

System Architecture:

1. Sensing Layer: Raspberry Pi captures live video, processes frames with YOLOv5, and filters ambulance detections.

2. Signal Transmission Layer: Detected ambulances trigger alerts via MQTT and LoRa communication to traffic controllers.

3. Action & Control Layer: Traffic signals are preempted to clear a path for ambulances, with visual and auditory confirmation.

Software & Hardware:

• Python, PyTorch, YOLOv5, ONNX Runtime, OpenCV, LabelImg, and MQTT for software; Raspberry Pi 3 Model B+ for hardware.

Conclusion

This paper presented a real-time ambulance detection system optimized for deployment on resource-constrained edge devices. By leveraging YOLOv5 and ONNX Runtime, the proposed approach achieves reliable detection performance under deployment constraints while maintaining feasible real-time performance on a Raspberry Pi. The comparative evaluation demonstrates that ONNX-based optimization significantly improves inference speed over native PyTorch execution, validating its suitability for intelligent traffic management systems. Future work will explore quantization and hardware accelerators to further enhance performance.NAlthough the proposed system demonstrates reliable detection performance, its inference speed is limited by CPU-only execution on the Raspberry Pi. Additionally, the dataset size, while sufficient for proof-of-feasibility, may not fully represent all ambulance designs and regional variations. Future work will address these limitations through dataset expansion and hardware acceleration.

References

[1] Redmon, J., Farhadi, A. (2018). YOLOv3: An Incremental Improvement. arXiv preprint arXiv:1804.02767. (Foundational YOLO paper) [2] Bochkovskiy, A., Wang, C. Y., & Liao, H. Y. M. (2020). YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv preprint arXiv:2004.10934. (Another key YOLO paper, relevant for context). [3] Jocher, G., et al. (2020). Ultralytics YOLOv5. (Direct citation for the YOLOv5 framework). [4] Microsoft. (n.d.). ONNX Runtime. (Official reference for ONNX Runtime) [5] Perera, P., & Oza, H. (2020). Real-Time Object Detection on Edge Devices for Smart Traffic Management. Proceedings of the IEEE International Conference on Industrial Technology (ICIT). (Example of deploying object detection on edge devices for traffic) [6] Srivastava, S., et al. (2019). Emergency Vehicle Detection and Prioritization Using Deep Learning. International Conference on Computer Vision and Image Processing (CVIP). (Directly relevant to emergency vehicle detection using deep learning) [7] Khan, R. U., & Abdullah, S. (2021). Efficient Object Detection for Resource-Constrained Embedded Systems. Journal of Network and Computer Applications. (Focuses on optimization for edge computing) [8] Raspberry Pi Foundation. (n.d.). Raspberry Pi OS. Official reference for the Raspberry Pi OS) [9] Wang, H., & Chen, J. (2022). ZEnhancing Traffic Management with YOLOv5-Based Ambulance Tracking System. Journal of Smart Cities and Society. (Example paper specifically on YOLOv5 for ambulance tracking or similar emergency vehicles) [10] Zhu, Y., et al. (2018). Deep Learning for Siren Detection and Localization in Urban Environments. IEEE Transactions on Intelligent Transportation Systems. (Relevant for the \"acoustic detection\" part of your literature review) [11] OpenCV. (n.d.). OpenSource Computer Vision Library. (Reference for the OpenCV library) [12] The Open Neural Network Exchange (ONNX) Community. (n.d.). ONNX: Open Neural Network Exchange. (General reference for the ONNX format) [13] Lin, T. Y., et al. (2017). Feature Pyramid Networks for Object Detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). (This paper introduced Feature Pyramid Networks (FPNs), which are a core component of modern object detectors like YOLOv5, enabling multi-scale detection crucial for varying object sizes in traffic scenes). [14] Chen, Z., et al. (2021). Real-time Emergency Vehicle Detection Using Lightweight Deep Learning Models on Embedded Systems. Sensors. (Directly relevant to your project, focusing on lightweight models and embedded systems for emergency vehicles, which aligns with your Raspberry Pi deployment). [15] Intel OpenVINO Toolkit. (n.d.). OpenVINO™ Toolkit. (While you used ONNX Runtime, OpenVINO is a major competitor for edge inference optimization, and referencing it can show awareness of broader optimization techniques, especially if discussing future scope or alternative deployment options).

Copyright

Copyright © 2026 Abhishek Khode, T. Rishi Teja, B. Sai Charan Goud, Dhanush Goud, Tejaswi Vallabhapurapu. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.