Lower Complex CNN Model for Fire Detection in Surveillance Videos

Recently, advances in the field of deep learning encourages the use of convolutional neural networks (CNN) based methods to efficiently detect fire by analyzing videos captured by surveillance cameras. These methods try to derive standard features of fire frames. This improves accuracy substantially but the main problem is these methods need higher memory, large amounts of computational time. The proposed CNN model is less complex than the existing one which is inspired by AlexNet and GoogleNet architecture. The model is trained with the variety and challenging fire and nonfire Furg dataset from Kaggle to achieve great efficiency and accuracy. It consists of 4800 images. Out of them 2817 are of lire images and 1983 images are of non-fire class. Proposed Lower complex CNN model which is variant of AlexNet by removing two layers from actual AlexNet Model.For a fire detection and binary classification task, the lower complex CNN model achieves an accuracy of 0.93.