Source-linked AI summary
FireNet: A Specialized Lightweight Fire & Smoke Detection Model for Real-Time IoT Applications
Arpit Jadon, Mohd. Omama, Akshay Varshney, Mohammad Samar Ansari, Rishabh Sharma
TL;DR
Fire detection systems must balance detection performance, response time, false-trigger control, and portability to low-cost hardware. The paper proposes FireNet, a lightweight neural network trained from scratch and evaluated on standard and custom datasets. It reports real-time operation at up to 24 fps on Raspberry Pi alongside encouraging performance across both datasets.
Problem
Fire detection research faces a trade-off between model performance and size, while existing datasets may lack the diversity needed for realistic scenarios.
Method
The authors design FireNet from scratch, train it on diverse data, and deploy it within an IoT fire-detection unit using Raspberry Pi and visual smoke sensing.
Results
Up to 24 fps is achieved on Raspberry Pi 3B, with the model evaluated on a standard fire dataset and a self-made realistic test dataset.
Takeaways & Limitations
FireNet provides a lightweight embedded approach for real-time fire detection with separate fire and smoke identification and remote visual alerts.
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
The paper targets fire detection systems that can trigger falsely or detect too late, proposing a vision-based, lightweight approach for embedded IoT deployment. FireNet is trained from scratch, distinguishes fire from smoke, and supports remote alerts.
- Existing fire alarm systems can falsely trigger, while systems requiring sufficient fire initiation may detect too late and cause irreparable damage.
- Visual fire detection is presented as an alternative for improving robustness and reliability over existing systems.
- The authors develop FireNet from scratch and train it on a dataset compiled from multiple sources.
- The system is deployed on Raspberry Pi and differentiates fire from smoke using distinct sound alarms to reduce false triggering.
- The proposed unit is intended to support early warnings, fire-brigade notification, and automatic fire suppression triggering.
A. Contributions
The paper contributes a shallow FireNet model, diverse fire datasets, and a complete IoT fire-detection implementation. Together, these components target real-time embedded detection with reduced false and delayed triggering.
- A. Contributions: FireNet is a shallow neural network designed to perform real-time fire detection at a higher frame rate than previous deep-learning approaches.
- A. Contributions: The authors introduce a small, diverse training dataset combining images from multiple sources and a self-made dataset of videos recorded in challenging environments.
- A. Contributions: The complete IoT unit is designed to reduce false and delayed triggering while providing remote verification through real-time visual alert messages.
- The paper presents its dataset, proposed approach, IoT implementation, results, and effectiveness discussion in successive sections.
II. RELATED WORK
Prior fire-detection methods include handcrafted feature-based systems and deep CNN approaches. Handcrafted methods are computationally efficient but tedious to design, whereas deep models automate feature extraction at the cost of heavier deployment requirements.
- Hand-engineered approaches: Handcrafted approaches use motion, color, edges, optical flow, or multiple sensors to distinguish fire and smoke from non-fire objects.
- Hand-engineered approaches: Hand-engineered methods can run efficiently on Raspberry Pi but require manual feature extraction, which becomes tedious and inefficient for large image datasets.
- Deep-learning approaches: Deep-learning approaches automate feature extraction and can be more efficient and reliable than conventional handcrafted image processing.
- Deep-learning approaches: Deep-learning fire detectors require substantial computational power during training and deployment, limiting their suitability for low-cost hardware.
- Deep-learning approaches: Existing CNN studies using architectures such as AlexNet, VGG16, ResNet50, SqueezeNet, GoogleNet, and MobileNetV2 face dataset-specific or model-size constraints.
III. DATASET DESCRIPTION
The authors identify insufficient diversity in existing fire datasets and construct datasets intended to better represent realistic detection conditions. Their test data combines fire videos, non-fire videos, and challenging non-fire images, and the resources are open-sourced.
- Dataset motivation: Existing fire datasets are scarce in diversity, with many similar images limiting expected performance in realistic fire-detection scenarios.
- Training dataset: The authors build a diverse training dataset from sampled existing datasets, internet images, and fire-like non-fire scenes.
- Test dataset: Realistic images are reserved for testing because the fire-detection unit must ultimately operate under those conditions.
- Test dataset: The test dataset contains 46 fire videos, 16 non-fire videos, and 160 challenging non-fire images, with randomly sampled frames forming the final set.
- The authors open-source the dataset and FireNet for research-community use and extension.
IV. PROPOSED APPROACH
FireNet addresses the deployment challenge posed by bulky CNN-based fire detectors by proposing a lightweight architecture for real-time operation on low-cost hardware such as Raspberry Pi.
- Bulky fine-tuned CNNs can have large on-disk models and many layers, limiting sufficiently fast real-time operation on low-cost Raspberry Pi hardware.
- Low-cost commercially available hardware is needed because the intended fire-detection units must be economically feasible for real-world environments.
- FireNet is designed for mobile and embedded applications, running at more than 24 frames per second on Raspberry Pi 3B.The network contains three convolutional layers and four dense layers, including the output softmax layer.
A. Architecture
FireNet uses a compact multilayer architecture that processes 64×64×3 color images through convolutional and dense layers to produce a two-neuron output prediction.
- FireNet has 14 total layers, including pooling, dropout, and softmax output layers, with 646,818 trainable parameters occupying approximately 7.45 MB on disk.
- The first convolutional layer accepts colored 64×64×3 images and uses 16 filters with 3×3 kernels.
- The two subsequent convolutional layers double input features while retaining the kernel size, followed by flattening and dense layers of 256 and 128 neurons.
- The final dense layer contains two neurons and serves as the output prediction layer.
B. Regularization
The network applies dropout in both convolutional and dense layers after observing improved overall results with convolutional-layer dropout.
- Dropout values are 0.5 for convolutional layers and 0.2 for the subsequent dense layer.The authors state that dropout in convolutional layers improved the network's overall results.
V. COMPLETE FIRE DETECTION UNIT AND IOT IMPLEMENTATION
The complete unit combines FireNet on Raspberry Pi 3B with smoke sensing, distinct alarms, and cloud-based messaging to distinguish fire from smoke and provide remote visual alerts.
- FireNet is deployed on Raspberry Pi 3B with a smoke sensor and two distinct fire alarms for separate fire and smoke detection.The distinct alarms are intended to eliminate false triggering associated with conventional smoke-based fire detectors.
- AWS S3 stores fire images or clips captured during an emergency, while Twilio sends an MMS containing the visual evidence and fire alert message.
- The microcontroller converts the smoke sensor's analog data for Raspberry Pi and triggers the sound alarms.
VI. RESULTS
FireNet was evaluated on a real-world dataset and a comparison dataset using multiple performance metrics, training curves, and a non-overlapping validation split. The model also runs at 24 frames per second on Raspberry Pi 3B.
- Datasets: The evaluation used real-world fire and non-fire images compiled from challenging environments, alongside the less diverse Foggia dataset for comparison.The real-world images were intended to reflect camera quality encountered on low-cost embedded hardware.
- Evaluation metrics: Four metrics—accuracy, precision, recall, and F-measure—were used to provide a complete performance analysis on both datasets.The reported metric values were described as satisfactory on both datasets.
- Training analysis: A 70% training and 30% validation split with no overlap between sets was used to assess generalization through training curves.The training and validation curves were presented as evidence of validation-set generalization.
- Embedded performance: 24 frames per second were achieved on Raspberry Pi 3B, supporting real-time execution on the embedded platform.The paper identifies this frame rate as a significant network advantage.
VII. DISCUSSION ON SIMILAR WORKS
The discussion positions FireNet as a compact alternative to larger CNN-based fire-detection models. Its design and diverse training data are associated with real-time performance on low-cost embedded hardware.
- Model size: FireNet occupies approximately 7.45 MB on disk and contains 646,818 trainable parameters because of its shallow network.The paper presents these properties as its main size advantage over larger CNN-based approaches.
- Comparison scope: FireNet’s comparison depends on resource constraints and dataset choice, since the paper notes that better-performing solutions also exist.The discussion defines “better” relative to the resources used and the dataset trained on.
- Real-time performance: Other massive CNN models typically report around 4-5 frames per second on low-cost embedded hardware, whereas FireNet provides up to 24 fps.The paper attributes FireNet’s capability to its diverse training dataset and specialized design from scratch.
VIII. CONCLUSION
The paper concludes that FireNet is a lightweight, diverse-data-trained neural network for real-time fire detection on Raspberry Pi-class hardware. Its evaluation reports encouraging performance across standard and self-made datasets, with IoT alerting and visual feedback.
- Contribution: FireNet is a very lightweight neural network built from scratch and trained on a diverse dataset.The model is intended for an IoT-capable fire-detection unit.
- Embedded deployment: 24 frames per second on Raspberry Pi 3B enables the proposed unit to operate in real time on low-cost embedded hardware.The paper reports performance in terms of accuracy, precision, recall, and F-measure on two datasets.
- System scope: The IoT functionality provides real-time visual feedback and fire alerts, while future work targets evaluation on an even more diverse dataset.The conclusion reports encouraging results on both a standard fire dataset and a self-made test dataset.