Fire Dispersal Estimation in Videos using Background Modelling and Subtraction by Tensor Decomposition

In this paper a comparative analysis is presented among 4 algorithms employing tensor representation of videos for background modelling and subtraction aiming the estimation of fire dispersal. The algorithms are HoRPCA by IALM, Tucker-ALS, CP-ALS, and t-SVD. They are applied over a database of 6 videos containing fires at different stage of spreading, recorded at different scale and angle of perspective. In part of the videos intense smoke is also present. Decomposition times, full processing times with preprocessing stage and accuracy of fire dispersal in terms of relative number of correctly detected pixels forming the flame areas to all flame pixels from the original recordings are registered. Positive results are obtained which reveal the applicability of the tested algorithms for fire dispersal estimation and with the in-depth analysis of experimental results a selection in order of preference could be made for future applications given the circumstances at which fire breaks out.