A Universal Foreground Segmentation Technique using Deep-Neural Network

Abstract

Background subtraction is generally used for foreground segmentation (moving object detection) from video sequences. Several background subtraction methods have been proposed for visual surveillance applications. However, the existing methods fail in case of real-surveillance challenges such as camouflage, sudden illumination variation, hard shadow, camera-jitter, non-static background, etc. A deep-neural network based background subtraction model is presented for flexible foreground segmentation. In addition to background subtraction model, a novel background modeling technique is also proposed for flexible background subtraction process. The presented deep-neural network architecture performs the background subtraction operation using the non-handcrafted features. The proposed method uses optical-flow details to make use of temporal information. This temporal information and spatial information (from the background image and current processing frame) are used for the training purpose. The model is trained using randomly selected images and its ground truth images from CDnet-2014 dataset. The presented model is evaluated using CDnet-2014 dataset, and it gives significant results compared to the existing background subtraction methods in terms of qualitative and quantitative analyzes.