Object Detection through Fires Using Violet Illumination Coupled with Deep Learning

Abstract

Fire accidents threaten public safety. One of the greatest challenges during fire rescue is that firefighters need to find objects as quickly as possible in an environment with strong flame luminosity and dense smoke. This paper reports an optical method, called violet illumination, coupled with deep learning, to significantly increase the effectiveness in searching for and identifying rescue targets during a fire. With a relatively simple optical system, broadband flame luminosity can be spectrally filtered out from the scattering signal of the object. The application of deep learning algorithms can further and significantly enhance the effectiveness of object search and identification. The work shows that this novel optics–deep learning combined method can improve the object identification accuracy from 7.0% with the naked eye to 83.1%. A processing speed of 10 frames per second can also be achieved on a single CPU. These results indicate that the optical method coupled with machine learning algorithms can potentially be a very useful technique for object searching in fire rescue, especially considering the emergence of low-cost, powerful, compact violet light sources and the rapid development of machine learning methods. Potential designs for practical systems are also discussed.