Wildfire detection system

Customer: Incom Group (Tomsk, Russia)

Summary: This study was launched in response to the widespread wildfires in Siberia's wild forests in 2019. Its primary goal is to detect and locate wildfires and then notify the forest protection service. The system is based on the use of deep learning techniques, specifically an object detection network trained to identify wildfire smoke and its location. To train the network, a variety of open-source datasets were collected from universities and organizations such as the Nevada Seismological Laboratory, the Center for Wildfire Research, and Perm Forestry. A total of 1000 videos were collected, including 766 fire and 234 non-fire recordings. Seven frames were extracted from each video for training, which was determined experimentally. The networks were trained and tested on 6300 and 700 images, respectively. Different object detectors were tested, including EfficientDet-D0, EfficientDet-D1, SSD ResNet50 V1, SSD MobileNet V2, Faster R-CNN ResNet50 V1, and Faster R-CNN Inception ResNet V2. The most efficient detector in terms of localization accuracy was found to be EfficientDet-D1, which achieved a classification accuracy of 81% and a localization accuracy (mAP) of 87%. The processing speed of the network using a GeForce RTX 2080 Ti was found to be fast enough at approximately 9 frames per second.

Collaborators: Olga Gerget (Tomsk Polytechnic University, Tomsk, Russia)

Project type: Commercial

Media: Visualization of the model testing