Affordable Autonomous Solutions
The system designed by the researchers consists of three main modules.
The first uses a type of image analysis neural network called Attention-Gated U-Net to focus on specific relevant parts of an image in a very detailed and structured way, allowing it to identify segmentations of smoke and suspicious flames and reduce the risk of false alarms.
The second module locates the UAV and the fire. The system relies on a monocular simultaneous localization and mapping (SLAM) algorithm, a technique used in computer vision to track and map the camera's environment. This is applied in conjunction with a wildfire site distance estimation algorithm that predicts the spread of wildfire embers using advanced models based on variables such as wind, ember size, terrain, and heat, and the UAV's own GPS system to pinpoint the location of suspected wildfire sites.
Finally, a transformation matrix is used to register the visible and infrared images, which helps to avoid false positives of forest fires and allows us to see the complete image without omitting important details or segments.
The researchers note that the visible light cameras they used in their study are common, affordable, and capable of capturing images of rising smoke. But because forest fires typically start below the tree canopy and cannot be detected by traditional visible light cameras, the infrared cameras can detect the radioactive emissions and heat signatures of new fires, allowing the system to pinpoint their precise geographic location before they escalate into raging conflagrations.
While field tests proved their method works on a small scale, Zhang and his co-authors say there's a lot more work to be done before their method can be adopted by firefighting agencies. Capacity and computer resources are obstacles they still have to overcome. But Zhang is confident their approach is heading in a promising direction.
“In a sense, our research is at the forefront of investigating and validating this technology, and developing theories for future practical applications.”
PhD student Linhan Qiao and master’s student Shun Li 23 are co-first authors. Jun Yan, associate professor and Concordia University Research Chair (Tier 2) in Artificial Intelligence in Cybersecurity and Resilience, also contributed to the study as Li’s co-supervisor.
This research was supported by the Natural Sciences and Engineering Research Council of Canada (NSERC) and the Quebec Aerospace Research and Innovation Consortium's “AirTanker Visual Intelligent Tracking & Airborne Guidance System (AVITAGS)” project.
Read the cited paper, “Early detection and distance estimation of wildfires using aerial visible and infrared imagery.”
