SBIR Phase I: AI Enabled Cameras for Fire Detection

点击翻译按钮获取中文摘要

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will result from the potential to reduce devastation from forest fires. Wildfires across the world increasingly cause devastating impacts on local economies, human life, air quality, natural resources, property value, jobs and wages. In 2018, more than 58 million wildfires burned 8.8 million acres in the U.S. A system that alerts authorities within minutes (vs. days) dramatically increases the ability to minimize property damage and loss of life. The goal of this project is to develop a precise computationally efficient AI-based software solution for thermal (infrared) imaging detection of fire, to be implemented on various remote and moving platforms, such as manned aircraft and unmanned aerial vehicles. The resulting technology will be more precise than current solutions due to the incorporation of a unique approach to machine learning based on the use of real-life and artificially created training datasets. In addition, it will be developed to function “at the edge” to enable local, real-time operations in remote locations. A Deep Convolutional Learning approach should dramatically improve current mathematical models, potentially saving thousands of lives and billions of dollars.This Small Business Innovation Research Phase I project will demonstrate how a solution utilizing Deep Convolutional Learning can be successfully applied to the detection of forest fires from moving or static platforms or in remote locations. The project takes a unique approach to image recognition using multiplication-free neural networks based only on the discriminator of a generative adversarial network (GAN) and training the model with real and artificially created datasets that address extensive variability of conditions. This will eliminate confusion in the detection of smoke plumes, which can be caused by clouds and reflections of bodies of water. Additionally, the development of low-power neural network algorithms for real-time processing of sensory data will enable the software to run on low-cost embedded computers for use in remote, resource-constrained environments. Expectations are for precision levels above 80%, false alarm rates at less than 1 per 24 hours, time to detection in minutes vs. days and ease of use in the most remote locations.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.