Faster R-CNN Based Indoor Flame Detection for Firefighting Robot

Abstract

Firefighters are primarily tasked to handle fire incidents, but they are often exposed to high risks when extinguishing fire. Firefighting robots are actively being researched to reduce fire fighters injuries and deaths as well as increase their effectiveness on performing tasks. However, the major concern is how to make the flame detection methods to satisfy the high precision requirement of firefighting robot. Therefore, the flame detection of firefighting robot has become a hot topic in this area. In this paper, a Faster R-CNN model is proposed to detect flame in noisy images of fire ground. Firstly, the region generation network is used to extract the candidate flame regions. Secondly, the candidate flame regions are convoluted and pooled to extract the flame characteristics. Thirdly, the output features of Region Proposal Network (RPN) are fed into two fully connected layers: a box-regression layer which recognizes the locations of objects and a box-classification layer which classifies the objects. The dataset used in the experiment was obtained by video capture. The network is pre-trained based on Google platform Tensorflow, and the obtained precision and frame rate of the proposed method are up to 99.8% and 1.4 FPS, respectively. The experimental results demonstrate that the method equipped merits such as automatically extract the flame characteristics, effectively improve the precision of flame detection, and has excellent generalization ability and robustness.