A deep neural network and rule-based technique for fire risk identification in video frames

Abstract

Automatically monitoring roadside fire risk plays a significant role in ensuring road safety by reducing potential hazards imposed to vehicle drivers and enabling effective roadside vegetation management. However, little work has been conducted in this field using video data collected by vehicle-mounted cameras. In this paper, a novel approach is proposed for roadside fire risk identification based on the biomass of grasses. Inspired by the biomass measurement method by human in grass curing, the proposed approach predicts the biomass and identifies high-risk regions using threshold-based rules based on two site-specific parameters of roadside grasses--brown grass coverage (BGC) and height (BGH). The BGC is calculated as the percentage of brown grass pixels in a sampling region, while the BGH is predicted based on the connectivity characteristics of grass stems along the vertical direction. To further reduce the false alarm rate of fire risk, we additionally incorporate and compare two deep learning techniques, including autoencoder and convolutional neural network, for refining the results. Our approach shows high performance of combining threshold-based rules with deep neural networks in classifying low and high fire risk on a roadside image dataset from video collected by the Department of Transport and Main Roads, Queensland, Australia.