On Spatiotemporal-Aware Deep Neural Networks for Real-Time Video Fire Detection: Empowering Image-Based Models with Temporal and Spatial Features of Video

Fire is one of the natural events that has various risks for humans and the environment. Therefore, reliable and early detection of fire could greatly help save lives and prevent disasters. Image-based fire detection methods can be very quick and reliable, unlike, conventional chemical-based fire detectors. Convolutional Neural networks (CNN) have shown relatively impressive accuracy in the field of fire detection in recent years. However, extracting temporal features from images by CNN Networks proves to be the main challenge. In order to introduce temporal features inside CNN, we used motion and color features to improve the accuracy of the CNN. We propose a modified Resnet50 architecture by adding temporal features with the help of a classical image processing algorithm. Our implementation and results show that our improved Resnet50 architecture outperforms original Resnet50 on test images. Our results demonstrate that the performance of Deep Neural Networks (DNNs) can be improved by employing additional temporal features. We use motion and color features to improve the DNNs model (spatiotemporal-aware) to classify fire images more accurately than basic original DNNs without additional temporal features (spatiotemporal-unaware).