Background subtraction by probabilistic modeling of patch features learned by deep autoencoders.

Abstract

Video sequence analysis systems must be able to operate for long periods of time and they must attempt that the different events that can affect the quality of the input data do not diminish the performance of the system to an excessive extent. In this work a method called Probabilistic Mixture of Deeply Autoencoded Patch Features (PMDAPF) is proposed. A Deep Autoencoder is the cornerstone of the methodology for robust background modeling and foreground detection that is presented in this document. Its purpose is to obtain a reduced set of significant features from each patch belonging to one of the several shifted tilings of the video frames. Then, a probabilistic model is responsible for determining whether the whole patch belongs to the background or not. Foreground pixel detection takes into account the information of all patches in which the pixel is included. The robustness of the proposal, as well as its suitability to the uninterrupted analysis and processing of visual information, is reflected in the experiments, in which the performance of the proposed system is affected slightly whereas those of the classic methods are degraded drastically.