Supervised Methods for Foreground Segmentation and Object Detection: A Review on the CDnet2014 Data

Abstract

Foreground segmentation and background subtraction are essential tasks in computer vision, particularly for detecting moving objects in videos. The CDNet2014 dataset is commonly used to evaluate segmentation algorithm performance across various scenarios, such as lighting changes, camera motion, and camouflage effects. This paper reviews recent deep learning-based approaches, including convolutional neural networks (CNNs), autoencoders, generative adversarial networks (GANs), and interactive methods, with a focus on top-performing techniques like FgSegNet, Motion U-Net, and BSUV-Net. We compare results on CDNet2014 in terms of standard metrics like F -measure, precision, and recall, and discuss the generalization capabilities of different methods to unseen videos. By identifying each approach's strengths and weaknesses, we suggest research directions to overcome persistent challenges in this field, especially by enhancing robustness to scene variations and developing lightweight models suitable for real-time applications.