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

Arpit Jadon, Mohd. Omama, Akshay Varshney, Mohammad Samar Ansari, Rishabh SharmaarXiv 2019 · OpenReview

TL;DR

This paper introduces FireNet, a novel, lightweight neural network designed from scratch for real-time fire and smoke detection on resource-constrained IoT devices like the Raspberry Pi. FireNet achieves a compelling balance between high detection accuracy and a small model size, enabling its deployment on embedded platforms without sacrificing performance.

What Problem Does This Paper Address?

Fire disasters pose a significant threat to life and property, making early and reliable detection crucial. Traditional fire alarm systems often rely on physical sensors (smoke, thermal, flame detectors) which can suffer from false triggers or delayed detection. While vision-based approaches using deep learning have shown promise, existing models are often too computationally intensive and large to be deployed on low-cost, embedded IoT devices. This creates a gap for efficient, accurate, and affordable real-time fire detection solutions suitable for widespread adoption.

Methods

The core of this work is the development of FireNet, a custom-designed, shallow convolutional neural network (CNN). Unlike many existing approaches that fine-tune large, pre-trained CNNs (like VGG16 or ResNet), FireNet is built from the ground up to be lightweight. Its architecture consists of three convolutional layers, followed by pooling and dropout for regularization, and then three dense layers culminating in a softmax output layer. This "designed-from-scratch" approach allows for fine-grained control over model complexity and size.

To further enhance robustness and reduce false alarms, FireNet is integrated into a complete IoT system. This system includes a smoke sensor to aid in differentiating smoke from fire, and distinct sound alarms for each. An IoT component leverages cloud services (AWS S3 and Twilio) to send real-time visual feedback (images/clips) and alerts to end-users via MMS.

Datasets

The authors created a new, diverse dataset for training and testing, addressing limitations in existing benchmarks.

  • Training Dataset: A custom dataset comprising 1,124 fire images and 1,301 non-fire images. This dataset was compiled by collecting images from various sources, including Flickr, Google, and existing datasets like Foggia's and Sharma et al.'s, with augmentation applied to enhance diversity.
  • Test Dataset: A more realistic test dataset was compiled from self-shot videos in challenging environments. It includes 46 fire videos (19,094 frames) and 16 non-fire videos (6,747 frames), along with 160 challenging non-fire images. For comparison, the model was also evaluated on the Foggia dataset.

Key Insights and Ideas

  • "Designed-from-scratch" for Efficiency: Building a shallow CNN like FireNet from the ground up, rather than fine-tuning larger models, is key to achieving a small model size and high frame rates on embedded hardware.
  • Dataset Diversity is Crucial: The authors highlight that existing datasets often lack diversity, leading to models that perform poorly in varied real-world scenarios. Their custom dataset, incorporating images from multiple sources and challenging environments, is vital for robust performance.
  • Hybrid Detection for Reliability: Combining visual detection (FireNet) with a physical smoke sensor significantly reduces false alarms and improves the overall reliability of the fire detection system.
  • IoT Integration for Actionable Alerts: The seamless integration with IoT services allows for immediate visual verification and alerts, empowering users to respond effectively to emergencies.

Key Results

  • Real-time Performance: FireNet achieves a frame rate of over 24 frames per second on a Raspberry Pi 3B, enabling real-time fire detection.
  • Accuracy: On their custom real-world test dataset, FireNet achieved an accuracy of 93.91%, with a recall of 94% and precision of 97%. On the Foggia dataset, performance was even higher, with 96.53% accuracy.
  • Model Size: The trained FireNet model has a small on-disk size of approximately 7.45 MB, with around 646,818 trainable parameters, making it suitable for embedded deployment.
  • Reduced False Triggers: The integration of a smoke sensor and distinct alarms helps to mitigate false triggering issues common in traditional systems.

Limitations and Future Work

  • Dataset Size: While diverse, the authors acknowledge that the dataset may appear small. They suggest that future work could involve expanding the dataset further.
  • Microcontroller Dependency: Currently, a microcontroller is used for analog-to-digital conversion of the smoke sensor data. The authors propose replacing this with a dedicated ADC for Raspberry Pi (e.g., MCP3008) in future iterations.
  • Performance on More Diverse Datasets: While performance is encouraging, the authors aim to improve the model's performance on even more diverse datasets in future work.

Summary updated

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.
BibTeX
@misc{jadon2019firenet,
  author = {Arpit Jadon and Mohd. Omama and Akshay Varshney and Mohammad Samar Ansari and Rishabh Sharma},
  title = {{FireNet: A Specialized Lightweight Fire \& Smoke Detection Model for Real-Time IoT Applications}},
  year = {2019},
  eprint = {1905.11922},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1905.11922},
}
Figure