[Paper Review] FireNet: A Specialized Lightweight Fire & Smoke Detection Model for Real-Time IoT Applications

Arpit Jadon, Mohd Omama+3 more|arXiv (Cornell University)|May 28, 2019

Fire Detection and Safety Systems17 citations

TL;DR

FireNet is a lightweight, custom-designed convolutional neural network for real-time fire and smoke detection on low-cost IoT devices like the Raspberry Pi 3B. Trained on a diverse, real-world dataset, it achieves 95% F-measure and 24 FPS inference, outperforming existing models in accuracy and efficiency while reducing false alarms through combined vision and sensor-based detection.

ABSTRACT

Fire disasters typically result in lot of loss to life and property. It is therefore imperative that precise, fast, and possibly portable solutions to detect fire be made readily available to the masses at reasonable prices. There have been several research attempts to design effective and appropriately priced fire detection systems with varying degrees of success. However, most of them demonstrate a trade-off between performance and model size (which decides the model's ability to be installed on portable devices). The work presented in this paper is an attempt to deal with both the performance and model size issues in one design. Toward that end, a `designed-from-scratch' neural network, named FireNet, is proposed which is worthy on both the counts: (i) it has better performance than existing counterparts, and (ii) it is lightweight enough to be deploy-able on embedded platforms like Raspberry Pi. Performance evaluations on a standard dataset, as well as our own newly introduced custom-compiled fire dataset, are extremely encouraging.

Motivation & Objective

• Develop a highly efficient, real-time fire and smoke detection system suitable for deployment on low-power embedded devices.
• Address the limitations of traditional sensor-based fire detectors, such as false positives and delayed detection.
• Create a custom, diverse real-world fire dataset to better reflect practical imaging conditions on low-cost cameras.
• Integrate vision-based deep learning with physical smoke sensors to improve detection reliability and reduce false alarms.
• Enable remote monitoring via IoT for immediate user alerts using MMS with visual feedback.

Proposed method

• Designed a lightweight, custom convolutional neural network (FireNet) from scratch with only 646,818 trainable parameters to ensure low computational overhead.
• Trained FireNet on a combination of a standard public dataset (Foggia et al.) and a newly compiled, diverse real-world dataset captured under challenging conditions.
• Employed a 70/30 train-validation split with no overlap to ensure generalization, validated using training and validation loss and accuracy curves.
• Integrated a physical smoke sensor alongside the vision model to enhance reliability and reduce false positives from visual-only detection.
• Deployed the model on Raspberry Pi 3B (1.2GHz CPU, 1GB RAM) to achieve real-time inference at 24 frames per second.
• Implemented an IoT-based remote alert system that sends MMS with visual fire feedback and alerts to users upon detection.

Experimental results

Research questions

• RQ1 Can a lightweight, custom-designed CNN achieve high accuracy in real-time fire and smoke detection on low-cost embedded hardware like Raspberry Pi 3B?
• RQ2 How does performance on a custom, real-world dataset compare to performance on standard benchmark datasets in terms of accuracy and robustness?
• RQ3 To what extent can combining vision-based deep learning with physical smoke sensors reduce false positive triggering in fire detection systems?
• RQ4 Can the proposed system maintain high frame rates (e.g., 24 FPS) while ensuring high detection accuracy and low model size?
• RQ5 How effective is the IoT-enabled remote alert system in delivering timely visual and textual fire alerts to end users?

Key findings

• FireNet achieved 93.91% accuracy, 97% precision, 94% recall, and 95% F-measure on the custom real-world test dataset.
• On the standard Foggia dataset, FireNet achieved 96.53% accuracy, 95.54% precision, 97.46% recall, and 96.49% F-measure, demonstrating strong generalization.
• The model successfully runs at 24 frames per second on Raspberry Pi 3B, enabling real-time detection suitable for live video streams.
• The training and validation curves show no significant overfitting, indicating strong generalization with minimal gap between training and validation performance.
• The integration of a physical smoke sensor significantly reduces false positives compared to vision-only systems, improving system reliability.
• The IoT-based remote alert system successfully transmits MMS with visual feedback and fire alerts, enabling timely user response.