Wildfire Detection

Wildfire Detection

Over the last few years, wildfires in the forests of Cyprus have increased dramatically at an average of 150 wildfires every year. To overcome this issue, we developed an object detection approach based on the state-of-the-art deep learning model YOLO version 7. The model was trained to detect early-stage smoke from any distance in an RGB camera video stream input captured by unmanned aerial vehicles. 

The unmanned aerial vehicles were manufactured by the Unmanned Systems Research Laboratory (USRL) of the Cyprus Institute (CyI), which also led the data collection. Videos were collected at two different locations with different types of terrain and environmental conditions, Orounta and Lythrodontas, to improve the model’s performance and generalizability and allow wildfire detection in any real-world scenario tailored to the Cypriot landscape. Videos were manually labeled by our group using an annotation tool. 

A pipeline was constructed to reduce the latency of real-time detection using our own proprietary software tool DgiStreamer. We incorporated in the pipeline a tracking algorithm to stabilize the bounding box detection across time and a message broker to send the results of the real-time detection to a server located at a different machine. 


Related Papers