The advent of deep learning networks has led to an improvement in almost every area in the computer vision field. In this work, a foreground detection method is proposed which intends to improve algorithms within video surveillance systems. Specifically, the proposed approach consists of a principled probabilistic model that combines both the output information of a semantic segmentation convolutional neural model and the color value for each pixel. The relevant features are transformed in a nonlinear way so as to enhance the performance of the probabilistic model. In order to determine the method feasibility, a set of experiments based on video sequences that belong to several public repositories have been carried out. The results show that the foreground detection performance of the proposal is higher than that of traditional algorithms in many situations.