FDAR – Fire Detection and Ranging

A modular vision-based system for real-time fire, smoke, and human detection with 3D localization, designed for autonomous firefighting drones. The pipeline integrates RGB and thermal sensing, object detection, monocular depth estimation, and safety-aware targeting to support autonomous fire suppression.

Features

• Multi-Modal Perception
  • Supports RGB-only, thermal-only.
• Real-Time Detection
  • Includes YOLO (v5–v12) and transformer-based detectors (RT-DETRv2, RF-DETR).
  • Multi-class detection of fire, smoke, and humans.
• Depth Estimation
  • Monocular depth prediction using Depth Anything V2 with ZoeDepth scaling.
  • 3D localization of all detections.
• Fire Targeting & Safety
  • Lightweight fire source localization inside bounding boxes.
  • Human proximity checks with safe-to-suppress flags.
  • Optional pseudo-3D bounding box visualization.

Performance

Tested on NVIDIA Jetson Orin Nano (JetPack 6, CUDA 12.2), live camera stream via ROS2 Humble.

Configuration	Inference Speed
Original sequential pipeline	~2 FPS
Multi-threaded + TensorRT (FP16)	~20 FPS (10–25 FPS range)

• Fire detection confidence threshold: 0.3–0.4 for clear flames (detections below 0.1 discarded)
• Depth estimation error: ±0.25m at 3m distance
• Reliable ranging range after polynomial correction: 3m – 5m
• ROS2 camera Workswell WIRIS Enterprise latency: ~500–700ms (reduced from >1s via resolution optimization)
• No latency on ZED2 Stereolabs

Requirements

• Python 3.8+
• PyTorch and TochVision Unique Build
• OpenCV
• NumPy
• Other dependencies listed in requirements.txt

To install the wheels of Pytorch and TorchVision

#Once you clone the repository, use 
git lfs pull 
#The wheels are too heavy for Git

Project in Jetson Orin Nano

Everything was tested and made on the Jetson Orin Nano The project is already inside (Last Test 24/02/2026) To launch the project :

First Connect the camera model Workswell Wiris Enterprise

source /opt/ros/humble/setup.bash
#Inside the ros_ws project folder
source install/setup.bash
ros2 run workswell_ros2_interface camera_node

In another terminal, to run the FDAR script:

cd Documents/Stage_Saxion
#For the protection of the project I used a venv
#Launch the venv
source venv venv_saxion_project/bin/bash
#Once the venv Launch
cd DroneFirefighting
#The Main folder project of my work on the Jetson
source /opt/ros/humble/setup.bash
python Ros_run.py

Download of the project (Made for the Jetson)

git clone <>
cd <folder_name>
pip install -r requirements.txt
pip install torch-2.3.0a0+git97ff6cf-cp310-cp310-linux_aarch64.whl
pip install torchvision-0.18.0a0+6043bc2-cp310-cp310-linux_aarch64.whl

Installation on Nvidia Jetson Orin Nano

Install ROS2 on the Jetson Orin Nano (Humble). Connect the RGB–thermal camera to the Jetson.

Start the camera through ROS2 (With the Workswell Camera):

source /opt/ros/humble/setup.bash
#Inside the ros project folder
source install/setup.bash
ros2 run workswell_ros2_interface camera_node

In another terminal, source ROS2 and run the FDAR script:

source /opt/ros/humble/setup.bash
python Ros_run.py

Configuration Options

You can modify the following parameters in Ros_run.py:

• Input:

  • CameraNode : In the subscriber you can change the topic (for image and focal)

• Model Settings:

  • yolo_model: YOLOv5–YOLOv12, RT-DETRv2, or RF-DETR
  • depth_model: Depth Anything v2 model size: "small", "base", "large" OR Zoedepth ['train', 'infer', 'eval']

• Detection Settings:

  • conf_threshold: Confidence threshold for object detection
  • iou_threshold: IoU threshold for NMS
  • classes: Filter by class, e.g., [0, 1, 2] for specific classes, None for all classes

• Feature Toggles:

  • TensorBoard: For post video analysis

Project Structure

FDAR/
│── Ros_run.py                # Main entry point
│── detection_model.py    # Multi-modal object detection logic
│── depth_model.py        # Monocular depth estimation + scaling
│── bbox3d_utils.py       # 3D bounding box and projection utilities
│── load_camera_params.py # Camera intrinsics & extrinsics loader
├── requirements.txt      # Dependencies
└── README.md             # This file


## How It Works

1. **Object Detection**: Detects fire, smoke, and humans from RGB, thermal, or fused RGT images.
2. **Depth Estimation**: Generates relative depth maps and scales them to metric depth.