LAUSR

Indoor fire sensing robot could become the best choice to detect irregularity features in a complex situation with help of nascent visual technologies. This article works on improvement computational time and accuracy of video fire detection (VFD) using on those robots. The profile of flame is first identified by flame luminance and chrominance using the proposed VFD network. A more efficient layer aggregation connection is designed to improve detection accuracy. The proposed feature pyramid layer combining ghost convolution and cross-connection improves the feature extraction ability. Improvements in reparameterized blocks with residual connections increase inference speed. The irrigation feature of fast burning fire flame in the incipient, growth, full-developed and decay stages are fully detected with analysis of the proposed visualized Heat Release Rate (vHRR) with spark, flicker, and other dynamic disturbances involved. Compared with high-performance methods, the average detection accuracy of this method on public datasets and self-built datasets is improved by 3.5% and 2.3%. Five indoor representative scenarios illustrate that our method achieves a good balance between the detection accuracy and speed of flame trends. In the future, the model could be deployed in lightweight, fast-moving indoor sensor robots for fire detection in hazardous environments.