.png){kind=small}
Fire and Smoke Detection
Problem
In 2017 were recorded more than 200,000 forest fires in Brazil, which adversely affect the flora, fauna and human health. Forest fires are also devastating for the Amazonian forests, which are the “lungs” of our planet.
Besides this the government spends a large amount of money to prevent and cope with natural disasters. Because fires are stealing up people’s dwellings, farms, and other relevant infrastructure. To solve this problem, we were approached by TERA Informática, which develops software, primarily in the forestry industry.
At the beginning, TERA Informática installed 32 PTZ cameras in Brazilian forests. In the data center, where the picture flowed from the video cameras, the attendants watched around the clock. If a forest fire has been detected, they called the fire brigade. However, manual observation did not bring the desired results. So, they asked us to automate the process.
Solution
Our engineers have created a fire detector based on data from customer’s cameras. The neural network was trained on 30,000 images. Half of them contained images of fires, and the second half did not.
The algorithm works by processing frames on the CPU. A full cycle of image processing from 32 cameras, which constantly rotate 360 °, takes 22 seconds.
The accuracy of the algorithm is 97%. For one day, the detector helped the system operators to detect 713 of the 735 forest fires.
The detector was designed using the YOLO v3 architecture.
Testimonials
“Their algorithms increased the software offering's value. Receptive and accommodating, the team fostered a smooth and efficient process. Hey Machine Learning communicated seamlessly and worked around time zone differences”
Theo Fernandes, a systems analyst at TERA Informática
