Early wildfire point perception methods

Outdoor experiment video

Preception Methods ooverview

In this work, we implemented several functions based on DJI M300 Drone and H20T camera for the early wildfire point preception applications:

• The wildfire point is firstly segmented by CNN-based networks to provide the semantic information for the other submodules in the framework.

• After the indoor calibration, the precise camera trajectory with correct scale is recoveried by ORB-SLAM2 and the drone platform navigation information. Then the depth is estimated.

• A model-based visible-infrared images registration is proposed to fuse the two types of information to reduce the false positive alram further.

Features

features
Attention gate U-net wildfire segmentation
Trianglulation-based wildfire point depth estimation
Visible-infrared camera system calibration
Model-based wide fire point registration

Realted Works

• We use the forest fire detection system to control M300 and capture the data during the trajectory.

Hardware

• DJI M300 RTK
• DJI H20T Camera
• Nvidia NX on board computer

Software

• ORB-SLAM2
• g2o
• OpenCV
• Pytorch

How to build

For each submodules:

1. cd <submodules>
2. mkdir -p build&&cd build
3. cmake cmake -DCMAKE_EXPORT_COMPILE_COMMANDS=1 -DCMAKE_BUILD_TYPE=Release ..

The executable file will be generated under <submodules>/bin directory.

Copyright

Copyright (C) 2022 Concordia NAVlab. All rights reserved.