Gathering and Processing Image Data
The drones are interconnected via an outdoor Wi-Fi network with a speed of 1775 Mbps with coverage of 200-300 m at the 5 GHZ band. The drones run MAVROS, part of the Robot Operating System (ROS). It facilitates communication between the onboard computer and flight controller as well as among drones, allowing for autonomous swarm operations.
As directed by the manager drone, worker drones fly in a synchronized manner and capture images at specific time intervals. The images are compiled and fed into the NeRF model, which outputs a point cloud. The researchers further process the point cloud to remove the background noise and segment the smoke plume in 3D. They do this using a combination of two machine learning models: Gaussian Naive Bayes and YOLOv8 (You Only Look Once).
The entire process is repeated for each time interval.
Related: Delivering High-Resolution Machine Vision for Enterprise Drone Applications
Taken together, the succession of reconstructed models captures a plume’s “evolution over time, showcasing its growth, directional shifts, and eventual dissipation,” they write.
Validating the 3D Reconstruction Model
But does the 3D reconstruction model work in the field? To answer that question, the researchers conducted two tests.
In the first test, designed to measure accuracy, they deployed the drones at an altitude of 10 m over a Ford F-150 pickup truck to snap pictures. The subsequent point cloud had an error rate of 1.8% with a standard deviation of about 0.98%.
In the second test, designed to measure effectiveness, they used two smoke generators to produce smoke plumes that extended upwards to about 10 m in height with widths of 1-10 m. Each worker drone completed a full circular circuit in 32 seconds and recorded 260 images at 8 fps. Each drone completed five data collection circuits. After gathering and compiling the images, the researchers completed the reconstruction process, confirming that the process captures changes in the plume’s dynamics over time.
