Fire Detection Using Deep Transfer Learning on Surveillance Videos
Abdul Bari, Tapas Saini, Anoop Kumar
Abstract
Traditional sensor-based fire detection systems cannot be alerted until the heat actually reach to the sensors. Therefore, it is evident to make a fast, robust and reliable system which can detect fire at an early stage. We propose a method that is able to detect fire by analyzing videos acquired by surveillance cameras. Recent development in Deep Learning has been proved to be highly effective in the field of computer vision. Transfer Learning (a deep learning methodology) has emerged to be extremely helpful for the applications with scarcity of training data. Using Transfer Learning we leveraged the Deep Learning models already trained on ImageNet Dataset. To solve the fire detection problem, we fine-tuned these models using our own curated dataset which consists of videos downloaded from the internet and our own recorded videos. Specifically, we chose pre trained InceptionV3 and MobileNetV2 models for transfer learning and in between comparison. We have also shown comparison between transfer learning and full model training on the same data set. We found that transfer learned models perform way better than fully trained models when trained on limited dataset. Our models also outperformed the state-of-the-art hand-crafted features-based fire detection systems.