Contextual background modeling using deep convolutional neural network

Abstract

Moving object detection is a crucial problem in computer vision. This affects the performance of the overall system in surveillance applications. In this paper, a Deep-Convolutional Neural Network with fully convolutional approach is proposed. Convolutional networks are powerful models to extract hierarchies of non-handcrafted features. The primary objective of the paper is to build an accurate foreground segmentation system with limited user interventions. The presented work focuses to build a fully convolutional network with skip architecture to identify moving objects in complex scenarios. The network is modeled as an end-to-end fully convolutional network, and the method contains a new hierarchical pooling layer to make use of global contextual information. The presented model utilizes a pre-trained VGG-19 Net model for the construction of Deep-Convolutional Neural Network (Deep-CNN) model. The fine and coarse features are fused using skip architecture to improve the feature representation. The qualitative and quantitative performance of the Deep-CNN architecture is tested on ChangeDetection.net-2014 dataset. The results produced by the Deep-CNN method were compared with the techniques in the recent literature. The Deep-CNN method outperforms the state-of-the-art methods without relying on any post-processing techniques.