Early Fire Detection

Patch-based fire detection system combining classical vision and deep learning for fast, robust early-stage fire localization on edge devices.

Problem

Early-stage fire detection in wide-area surveillance footage is challenging due to small flame regions, off-center appearance, and strict latency constraints on edge devices. While object detection models can localize fire directly, they are often too heavy, data-hungry, and difficult to deploy in real-time monitoring systems.

This project explores an alternative design: high-speed fire classification + spatial reasoning, rather than full-frame detection.

Project Structure

The work progressed through five structured phases:

• Phase 0 — Data preparation and system design
• Phase 1 — Initial patch-based fire classification experiments
• Phase 2 — Detailed evaluation and comparison with detection models
• Phase 3 — Training strategy and robustness analysis

Each phase incrementally addressed a concrete limitation observed in the previous stage.

Phase 0 — Data Preparation

To overcome limited annotated fire datasets, data was aggregated from multiple public sources, including MIVIA, FireNet, SKLFS, and Roboflow fire datasets.

Instead of training on full images:

• Images were divided into fixed-size patches
• Patches were labeled as fire / non-fire based on overlap with ground-truth regions
• Balanced sampling ensured equal representation of fire and non-fire patches

This design simplified training while enabling dense spatial coverage.

Phase 1 — Initial Fire Classification

A lightweight CNN-based binary fire classifier was trained on image patches.

Advantages of this approach:

• Small model size
• Fast inference
• Simple annotation pipeline
• Suitable for edge deployment

However, early experiments revealed a key limitation:

Fire patches not centered within the window receive significantly lower confidence scores.

Phase 2 — Detailed Analysis and Limitations

Further evaluation showed that:

• Sliding-window classification is computationally expensive
• Fixed grid partitioning reduces computation but introduces off-center bias
• Flames near patch boundaries are frequently misclassified

Comparisons with YOLO-based detectors confirmed:

• Detection models are robust but costly
• Classification models are efficient but spatially sensitive

This led to a critical insight:

Spatial alignment matters more than model capacity for early-stage fire detection.

Phase 3 — Fire Centering and Training Strategy

To address off-center degradation, a fire centering and expansion strategy was introduced.

GMM-based Fire Masking

• A Gaussian Mixture Model (GMM) was trained on fire pixel color distributions
• Fire likelihood maps were generated via probabilistic inference
• Binary fire masks were obtained using thresholding and morphological filtering

Fire Centering

Using the estimated fire mask:

• The centroid of the fire region was computed
• Image patches were spatially transformed via folding and mirroring
• Fire regions were moved toward the patch center and enlarged

This preprocessing step significantly improved classification confidence without increasing model complexity.

Phase 4 — Full Algorithm and Inference Pipeline

The final system integrates classical vision and deep learning into a unified pipeline:

1. Divide input frames into grid-based patches
2. Apply GMM-based fire likelihood estimation
3. Perform fire centering and expansion if likelihood exceeds a threshold
4. Classify centered patches using the trained fire classifier
5. Aggregate patch-level predictions into a spatial fire map
6. Trigger alarms based on spatial density and temporal persistence

This design maintains real-time performance while improving robustness to flame position and scale.

Results and Insights

• Consistent accuracy improvement across multiple backbones
• Strong gains for corner and edge flame cases
• Minimal computational overhead per frame
• Competitive or superior performance compared to detection-based methods in early-stage fire scenarios

Notes and Lessons Learned

• Fire detection is a spatio-temporal reasoning problem, not purely a classification task
• Classical vision remains valuable when integrated structurally
• Spatial normalization can outperform architectural complexity
• Hybrid pipelines are often preferable for safety-critical, real-time systems

This project directly influenced later work on guided inference, global–local reasoning, and feedback-driven vision systems.