Wildfire detection is an active and challenging research topic in computer vision. Wildfires can cause damage to valuable natural resources and can harm people and their communities. In this paper, we present CICLOPE, a telesurveillance system that can perform remote monitoring and automatic fire and smoke detection. The CICLOPE system currently covers about 1.300,000 hectares of mainland Portugal enabling the monitoring of large areas at a very low cost per hectare. The goal of this paper is to present an assessment and evaluation of three of CICLOPE’s eleven automatic fire and smoke detection algorithms. Concretely, we aim to determine the potential benefits of three CICLOPE’s legacy algorithms: ADBACK, BEFORT, and BESTEST. ADBACK uses background subtraction techniques and a quasi-connected components method to detect smoke and fire while BEFORT and BESTEST use active learning techniques to update an auxiliary database of non-fire occurrences. We compare these three algorithms against one approach from the literature as a means to draw comparisons between the existing techniques. We also propose a number of performance metrics important to measure in a system of this kind, focusing on the consistent detection of smoke plumes and fire incidents over time, while achieving a low false positive rate. Using these evaluation measures, we show that our proprietary algorithms can attain the best performance on a dataset of 75 real wildfires over 3 months of surveillance.

This work has been partially supported by Portuguese National funds through FITEC - Programa Interface, with reference CIT "INOV - INESC Inovação - Financiamento Base".

Computer vision,Smoke and fire detection,Surveillance systems

• Ciências da Computação e da Informação - Ciências Naturais