A hybrid AI-human labeling framework with distributed edge-cloud architecture for accurate, cost-effective fire disaster management systems.
Real-time image capture at 10 FPS
JPEG compression & WiFi transmission
EfficientNetB3 + Bidirectional LSTM
Real-time severity classification
Advancing the state-of-the-art in vision-based fire detection through integrated innovations
Gemini 2.5 Flash automated annotation with 100% expert verification, achieving gold-standard dataset quality while reducing annotation time.
EfficientNetB3 for spatial feature extraction combined with bidirectional LSTM layers for temporal modeling of flame evolution patterns.
Distributed architecture using ESP-32 CAM modules for scalable, cost-effective real-time fire surveillance with sub-500ms latency.
From video capture to real-time severity classification
Granular severity assessment for informed emergency response decisions
State-of-the-art results on the MIVIA Fire benchmark dataset
Upload an image to see the fire severity classification in action
Drag & drop or click to select an image
Performance comparison with existing fire detection methods
| Method | Year | Approach | Accuracy | F1-Score |
|---|---|---|---|---|
| Chen et al. | 2004 | Color + Motion | 78.42% | 78.0% |
| Töreyin et al. | 2006 | Wavelet Analysis | 82.15% | 81.9% |
| BoWFire | 2015 | Bag-of-Visual-Words | 83.33% | 74-79% |
| Foggia et al. | 2015 | Multi-expert Fusion | 87.71% | 87.1% |
| ATT Squeeze U-Net | 2021 | Deep Learning | 89.45% | 88.9% |
| Two-Stream Net | 2024 | Spatial + Optical Flow | 92.15% | 91.5% |
| Proposed (Ours) NEW | 2024 | CNN-LSTM Hybrid | 94.39% | 93.78% |
Cost-effective solution for fire surveillance at scale
.png)