Scholarlink

<p>This paper introduces a novel foreground detection algorithm based on feature fusion convolutional neural networks (FFNet). Traditional foreground detection methods often struggle when dealing with complex backgrounds or when foreground objects are visually indistinguishable from the background. To address these challenges, we propose a method that first generates a Foreground Probability Map (FPM) using Robust Principal Component Analysis (RPCA), which effectively filters out background information while preserving foreground details. Subsequently, FFNet integrates the FPM with the inherent features of the original video frames through a feature fusion network. This fusion process enables FFNet to discern distinct foreground features, significantly enhancing detection accuracy. An adaptive parameter is also introduced to dynamically merge the FPM features with those of the original video frames, allowing for more nuanced feature integration. Experimental results on the CDNet2014 dataset demonstrate that our algorithm outperforms comparative methods, achieving superior F-measure values, particularly in scenarios with complex backgrounds or visually similar foreground and background objects. The project page is available at <a href="https://github.com/Lyoyot/FFNet.">https://github.com/Lyoyot/FFNet.</a></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Foreground detection through feature fusion convolutional neural networks: enhancing robustness against complex backgrounds

  • 期刊论文

  • 发表日期：

    2025年5月27日

查看全文
  • Youyou Lu,
  • Lixia Chen,
  • Xuewen Wang,
  • Xin Xu,
  • Xiaoli Sun

摘要

This paper introduces a novel foreground detection algorithm based on feature fusion convolutional neural networks (FFNet). Traditional foreground detection methods often struggle when dealing with complex backgrounds or when foreground objects are visually indistinguishable from the background. To address these challenges, we propose a method that first generates a Foreground Probability Map (FPM) using Robust Principal Component Analysis (RPCA), which effectively filters out background information while preserving foreground details. Subsequently, FFNet integrates the FPM with the inherent features of the original video frames through a feature fusion network. This fusion process enables FFNet to discern distinct foreground features, significantly enhancing detection accuracy. An adaptive parameter is also introduced to dynamically merge the FPM features with those of the original video frames, allowing for more nuanced feature integration. Experimental results on the CDNet2014 dataset demonstrate that our algorithm outperforms comparative methods, achieving superior F-measure values, particularly in scenarios with complex backgrounds or visually similar foreground and background objects. The project page is available at https://github.com/Lyoyot/FFNet.