A Novel Foreground Segmentation Method Using Convolutional Neural Network

Background subtraction is a commonly used approach for foreground segmentation (moving object detection). Different methods have been proposed based on this background subtraction technique. However, the algorithms give the false alarm in case of complex scenarios such as dynamic background, camera motion, shadow, illumination variation, camouflage, etc. A foreground segmentation system using convolutional neural network framework is proposed in this paper to handle these complex scenarios. In this approach, the non-handcrafted features learned from the deep neural network are used for the detection of moving objects. These non-handcrafted features are robust and efficient compared to the handcrafted features. The presented method is learned using spatial and temporal information. Additionally, a new background model is proposed to estimate the temporal information. We train the model end-to-end using input images, background images, and optical flow images. For the training purpose, we have randomly selected few images and its ground truth images from CDnet 2014. The proposed method is evaluated with benchmark datasets, and it outperforms the state-of-the-art methods in terms of qualitative and quantitative analyzes. The proposed model is capable of real-time processing because of its network architecture. Hence the model can be used in real-surveillance applications.