Realtime Pipeline Fire & Smoke Detection Using a Lightweight CNN Model

Fire disasters due to pipeline leaks result in loss to life and property. Therefore, an expeditious model to detect smoke and fire is much needed. Even though there have been several types of research done on fire and smoke detection, most of these focuses on very generic datasets that boast good performance; the case of monitoring pipelines remotely necessitates a focused approach for enhanced performance for early detection. This paper attempts to develop a model specifically for this purpose and can be deployed more confidently in a pipeline environment. We have also customized the existing dataset by adding images of pipelines to bias the dataset and give a more confidence rate while predicting. Our proposed neural network architecture achieves higher accuracy, precision, recall, and F-measure. Also, the lightweight makes it easily deployable on embedded platforms as well. The performance of our model is evaluated against FireNet on our biased dataset, sounds very promising.