Intelligent Spatial Estimation for Fire Hazards in Engineering Sites: An Enhanced YOLOv8-Powered Proximity Analysis Framework

This paper presents an enhanced dual-model YOLOv8 framework that integrates fire and smoke detection with proximity-based risk assessment to generate quantitative hazard scores and real-time situational awareness for engineering sites.

The Gist

Imagine you are the security guard for a massive, busy construction site. In the past, your job was simple: if you saw smoke, you yelled "Fire!" and ran to the nearest exit. But that's not enough anymore. You need to know: Is the fire just a small campfire in the distance, or is it a raging inferno about to swallow the workers and the fuel trucks?

This paper introduces a new "Super-Security Guard" powered by Artificial Intelligence (AI) that doesn't just see fire; it understands how close the danger is to the people and things that matter.

Here is the breakdown of how this "Super-Security Guard" works, using simple analogies:

1. The Two-Brain System (The Dual-Model)

Most old fire alarms are like a single person with a magnifying glass looking for smoke. This new system uses two specialized brains working together:

• Brain A (The Fire Expert): This is a highly trained AI (called YOLOv8) that has studied thousands of pictures of fires and smoke. Its only job is to spot the fire and draw a tight, precise outline around it, like a surgeon tracing the exact shape of a wound.
• Brain B (The World Observer): This is a general AI that has seen millions of everyday objects (cars, people, buildings, trees). It looks at the same scene and says, "Hey, I see a worker here, a truck there, and a gas tank over there."

The Magic: By combining these two brains, the system doesn't just know where the fire is; it knows who and what is standing next to it.

2. The "Ruler" Trick (From Pixels to Meters)

Cameras see the world in tiny dots called "pixels." If a fire is 50 pixels away from a worker on the screen, how far is that in real life? Is it 5 feet or 50 feet?

The researchers invented a clever trick to solve this. Imagine you know the worker in the video is exactly 6 feet tall. The system measures how many pixels tall the worker is on the screen. It uses that ratio as a ruler.

• Analogy: It's like looking at a map. If you know 1 inch on the map equals 1 mile in real life, you can measure the distance between two cities. This system does the same thing instantly, converting the "pixel distance" between the fire and the worker into a real-world measurement (like "3 meters away").

3. The "Danger Score" (Risk Assessment)

Once the system knows the fire is 3 meters away from a worker, it calculates a Risk Score. Think of this like a weather forecast, but for danger.

• Low Risk: A small fire 50 meters away from an empty field. The system says, "Keep an eye on it, but no panic."
• Critical Risk: A large fire 2 meters away from a fuel truck or a group of people. The system screams, "DANGER! Evacuate immediately!"

It calculates this score by looking at three things:

1. How big and confident is the fire? (Is it a spark or a blaze?)
2. Who is nearby? (Is it a person? A fuel tank? Or just a rock?)
3. How close are they? (The closer the fire, the higher the score).

4. The Visual Dashboard

Instead of just sending a text alert, the system draws on the video feed in real-time:

• It draws red lines connecting the fire to nearby objects.
• It writes the distance (e.g., "2.5m") right on the screen.
• It changes colors based on the risk level (Green for safe, Red for critical).

This allows a human manager to look at a screen and instantly understand the situation without doing any math. "Oh, that fire is close to the gas tank, but that one over there is far from everyone."

Why This Matters

In the past, fire detection systems were like a smoke detector in your kitchen: they tell you that there is smoke, but they don't tell you how bad it is or who needs to run.

This new framework is like having a smart, proactive safety officer who constantly asks: "The fire is here. The workers are there. They are getting too close. We need to act NOW."

It turns a simple "Fire!" alarm into a smart decision-making tool that helps engineers and safety teams prioritize their response, saving lives and protecting expensive equipment before the fire gets out of control.

Technical Summary

Here is a detailed technical summary of the paper "Intelligent Spatial Estimation for Fire Hazards in Engineering Sites: An Enhanced YOLOv8-Powered Proximity Analysis Framework."

1. Problem Statement

Traditional fire detection systems, whether sensor-based (smoke/thermal) or early vision-based approaches, suffer from significant limitations:

• Lack of Context: Most deep learning models focus solely on detecting the presence of fire or smoke (classification/detection) without analyzing the spatial relationship between the hazard and surrounding assets.
• Absence of Quantitative Risk: Existing systems rarely translate pixel-level detections into real-world physical measurements (e.g., meters), making it difficult to assess the immediate threat level to personnel, machinery, or infrastructure.
• Operational Gap: In safety-critical environments, risk is determined not just by the existence of fire, but by its proximity to vulnerable entities. Current frameworks lack an integrated pipeline that combines high-accuracy segmentation with real-time distance estimation and contextual risk quantification.

2. Methodology

The authors propose a Dual-Model Framework built on the YOLOv8 architecture, executed within a Google Colab environment using an NVIDIA L4 GPU. The pipeline consists of three main stages:

A. Data Preparation

• Dataset: A diverse dataset of 9,860 annotated images covering residential, forest, and industrial environments.
• Split: 78% (7,710 images) for training and 22% (2,150 images) for validation.
• Preprocessing: Images resized to 640×640 pixels with augmentations (flipping, scaling, mosaic) to improve robustness.

B. Dual-Model Architecture

The system processes each frame sequentially using two specialized models:

1. Primary Model (Fire/Smoke Segmentation): A custom-trained YOLOv8 instance segmentation model designed to detect and segment fire and smoke regions with high precision.
2. Secondary Model (Contextual Object Detection): A YOLOv8 model pre-trained on the COCO dataset to identify surrounding entities (people, vehicles, infrastructure, equipment).

C. Proximity Estimation & Risk Calculation

• Centroid Extraction: The geometric center (centroid) of the fire segmentation mask and the bounding boxes of surrounding objects are calculated.
• Distance Computation: Pairwise Euclidean distances are calculated in pixel coordinates between fire centroids and object centroids.
• Pixel-to-Meter Conversion: A scaling factor ($\kappa$) derived from known object dimensions or scene calibration converts pixel distances into approximate metric units (meters).
  • Note: The authors acknowledge these are approximations due to monocular projection limitations (lens distortion, depth variation) but argue they are sufficient for risk prioritization.
• Risk Assessment Model: A quantitative risk score ($R$) is computed for each fire-object pair using the formula:
  $$R_{ij} = H_i \cdot V(c_j) \cdot C_j \cdot E(d^m_{ij})$$
  Where:
  • $H_i$: Fire severity (based on segmentation confidence and spatial extent).
  • $V(c_j)$: Vulnerability weight of the object class (e.g., "person" > "wall").
  • $C_j$: Confidence score of the object detection.
  • $E(d^m_{ij})$: Exposure function, an exponential decay based on the metric distance ($d^m$).
• Alert Logic: The system aggregates risks to determine a frame-level score, categorizing threats into tiers (Low, Medium, High, Critical) and triggering alerts based on distance thresholds and risk levels.

3. Key Contributions

1. Dual-Model Pipeline: Integration of a specialized fire segmentation model with a general object detector to enable contextual scene understanding.
2. Spatial Risk Quantification: A novel mechanism to translate visual detections into interpretable real-world distances (meters) and quantitative risk scores, moving beyond simple "fire detected" alerts.
3. Actionable Visualization: The system generates annotated frames and animated sequences that overlay fire locations, detected objects, proximity lines, and distance values, facilitating intuitive situational awareness.
4. Open-Source & Lightweight: The framework is implemented using open-source tools, making it accessible and deployable in resource-constrained industrial settings.

4. Experimental Results

The model was trained for 200 epochs and evaluated on the validation set:

• Detection Performance:
  • mAP@0.5: > 91% (specifically 0.909–0.912 across runs).
  • Precision: > 90% (peaking at 91.2%).
  • Recall: ~82.5% (optimized for sensitivity to minimize false negatives in safety contexts).
  • F1 Score: 0.86 (at an optimal confidence threshold of ~0.31).
• Confusion Matrix: The model correctly classified 86% of true fire instances, with a bias toward sensitivity (prioritizing the detection of potential fires over perfect precision).
• Visual Output: The system successfully generated annotated animations showing fire contours, object labels, and red proximity lines with distance values, validating the end-to-end pipeline.

5. Significance and Impact

• Shift from Detection to Prioritization: This work transforms fire monitoring from a passive detection tool into an active risk management system. It answers not just "Is there fire?" but "How dangerous is this fire to specific assets?"
• Industrial Applicability: By providing distance-based risk scores, the system supports automated alerting and decision-making in engineering sites, allowing for targeted evacuation or resource deployment.
• Scalability: The use of standard cameras and open-source deep learning models makes the solution scalable for under-resourced environments where specialized 3D sensors (LiDAR) are unavailable.
• Future Integration: The framework provides a foundational layer for integrating with predictive analytics and automated emergency response systems.

In conclusion, the paper presents a robust, interpretable, and practical solution that bridges the gap between deep learning-based fire detection and real-world spatial risk assessment, offering a significant advancement in industrial safety surveillance.