Project Overview
Early detection of smoke and fire depends on quickly identifying where smoke originates. This project builds an end-to-end 2-class object detection system that simultaneously detects smoke regions and their origin points in still images and video feeds using YOLO object detection models.
Trained on a custom labeled dataset sourced and iterated through Roboflow Universe, the system was refined across 6 training experiments comparing YOLOv11n and YOLOv8n architectures on progressively improved dataset versions. The best model (YOLOv11n on dataset v6) achieved mAP50 of 0.679 and mAP50-95 of 0.347, demonstrating that dataset quality improvements consistently outperformed architecture changes alone.
Trained on a custom labeled dataset sourced and iterated through Roboflow Universe, the system was refined across 6 training experiments comparing YOLOv11n and YOLOv8n architectures on progressively improved dataset versions. The best model (YOLOv11n on dataset v6) achieved mAP50 of 0.679 and mAP50-95 of 0.347, demonstrating that dataset quality improvements consistently outperformed architecture changes alone.
The Challenge
Manual monitoring of surveillance or drone footage for smoke and fire is slow, fatiguing, and unreliable at scale. Human operators cannot continuously watch multiple camera feeds and early detection failures allow fires to spread before intervention. Existing general-purpose object detectors are not trained to distinguish smoke origin points from general smoke regions, making it difficult to localise the fire source for targeted response.
The Solution
Framed smoke origin detection as a 2-class object detection problem (smoke and origin) and built a full YOLO fine-tuning pipeline. A labeled dataset was sourced from Roboflow Universe and iteratively improved across 6 versions through augmentation tuning, preprocessing adjustments, and additional data collection.
Six training experiments were conducted comparing YOLOv11n and YOLOv8n on different dataset versions with varying augmentation strategies. Models were trained on an NVIDIA GeForce RTX 4080 Laptop GPU (CUDA 12.1) with image size 640, batch sizes 4–16, and early stopping patience of 30. Inference was tested on both images and video to validate real-world usability.
Six training experiments were conducted comparing YOLOv11n and YOLOv8n on different dataset versions with varying augmentation strategies. Models were trained on an NVIDIA GeForce RTX 4080 Laptop GPU (CUDA 12.1) with image size 640, batch sizes 4–16, and early stopping patience of 30. Inference was tested on both images and video to validate real-world usability.
Key Features
- End-to-end YOLO fine-tuning pipeline for 2-class smoke and origin detection
- 6 training experiments comparing YOLOv11n and YOLOv8n architectures
- Custom Roboflow dataset with 6 iterative versions improving augmentation and data quality
- Best model: YOLOv11n on dataset v6, mAP50 0.679, mAP50-95 0.347
- Per-class results: smoke recall 0.769, origin recall 0.523
- Evaluation metrics: Precision, Recall, mAP50, mAP50-95
- Inference on both still images and video feeds
- Training outputs saved per experiment for reproducibility
- NVIDIA RTX 4080 GPU training with CUDA 12.1 acceleration
- Roboflow API integration for dataset versioning and download
- Full pipeline in Jupyter notebook: setup, training, evaluation, inference
Results & Impact
Best performing model (YOLOv11n Exp 6) achieved mAP50 of 0.679 and mAP50-95 of 0.347. Smoke class shows strong recall of 0.769, meaning the model reliably detects visible smoke. Origin localisation (recall 0.523) is more challenging due to the smaller and less visually distinct nature of origin points.
Key finding: dataset quality improvements (more images, better versioning, safer augmentation) consistently delivered stronger results than architecture changes. YOLOv11n outperformed YOLOv8n on identical data (Exp 5 vs Exp 4). Further gains are expected from a larger, more diverse dataset. The system is deployable for real-time surveillance monitoring with further fine-tuning on domain-specific footage.
Key finding: dataset quality improvements (more images, better versioning, safer augmentation) consistently delivered stronger results than architecture changes. YOLOv11n outperformed YOLOv8n on identical data (Exp 5 vs Exp 4). Further gains are expected from a larger, more diverse dataset. The system is deployable for real-time surveillance monitoring with further fine-tuning on domain-specific footage.