Developing a neural network-based hardware and software system for early video detection of fires in real time

This article presents a method for early fire detection using video images from surveillance cameras, based on a temporal analysis of the suspicious area. The method makes it possible to compress the area of the video image to a time series, which is classified by a recurrent neural network. The time series contains metrics measured in the area and its surroundings to consider for its flicker. The flicker analysis determines whether the area is on fire. The proposed algorithm significantly reduces the number of false fire alarms as the video image are analyzed over time. The source data compression to a time series containing the area and surrounding characteristics allows using a small recurrent neural network to classify a suspicious area regardless of its size. These features make this model applicable for building an autonomous fire detector based on a single-board computer and a video camera. The paper describes the proposed model and the neural network training, training quality evaluation and the results of experiments, as well the results generated with the Jetson Nano single-board computer from NVIDIA.