Scholarlink

<p>Violence detection (VD) is a critical task in computer vision with applications in security, surveillance, and public safety. The proliferation of surveillance cameras and the increasing need for automated monitoring systems have underscored the importance of accurate and efficient violence detection algorithms. However, existing methods face several challenges, including limited performance in complex real-world scenarios, high false positive rates, and difficulties in capturing subtle violent behaviours. Addressing these challenges requires the development of advanced algorithms that can effectively differentiate between violent and non-violent activities while minimizing false positives. This paper proposes a Spatial–Temporal Context Collaboration Learning Multi-Stream Network (STCCLM-net) to address challenges in violence detection within video data, particularly in scenarios with crowded scenes, rapid motions, and occlusions. This framework leverages a dual-stream architecture comprising a spatial context extractor network (SCENet) and a temporal context extractor network (TCENet), incorporating a Spatial–Temporal Collaboration Unit (STCU) to optimize spatial and temporal features. The pre-processing stage involves frame difference and background suppression for motion capture and clutter removal. Feature extraction utilizes the VGG16 network for spatial and temporal context extraction, enhancing feature recognition. In this, fully connected VGG16 network (FC-VGG16 Net) efficiently handles sequential data with a one-dimensional structure, while convolutional VGG16 network (Conv-VGG16 Net) effectively models spatial–temporal sequences by leveraging convolution operations, enhancing the comprehensive modeling of spatial–temporal dynamics within video clip data. The STCU module explores the complementary nature of spatial and temporal features through an alternating co-attention mechanism, optimizing feature fusion. The classification stage labels video clips as violent or non-violent based on the extracted features. Together, these modules optimize the model's ability to discern complex patterns and salient features across spatial and temporal dimensions in video data. Experimental results showcase enhanced performance metrics, including increased precision and recall rates of overall 99.6%, validating its effectiveness in accurately classifying violent and non-violent actions on four video datasets.</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

A multi-stream framework using spatial–temporal collaboration learning networks for violence and non-violence classification in complex video environments

  • 期刊论文

  • 发表日期：

    2025年1月16日

查看全文
  • Barun Pandey,
  • Upasana Sinha,
  • Kapil Kumar Nagwanshi

摘要

Violence detection (VD) is a critical task in computer vision with applications in security, surveillance, and public safety. The proliferation of surveillance cameras and the increasing need for automated monitoring systems have underscored the importance of accurate and efficient violence detection algorithms. However, existing methods face several challenges, including limited performance in complex real-world scenarios, high false positive rates, and difficulties in capturing subtle violent behaviours. Addressing these challenges requires the development of advanced algorithms that can effectively differentiate between violent and non-violent activities while minimizing false positives. This paper proposes a Spatial–Temporal Context Collaboration Learning Multi-Stream Network (STCCLM-net) to address challenges in violence detection within video data, particularly in scenarios with crowded scenes, rapid motions, and occlusions. This framework leverages a dual-stream architecture comprising a spatial context extractor network (SCENet) and a temporal context extractor network (TCENet), incorporating a Spatial–Temporal Collaboration Unit (STCU) to optimize spatial and temporal features. The pre-processing stage involves frame difference and background suppression for motion capture and clutter removal. Feature extraction utilizes the VGG16 network for spatial and temporal context extraction, enhancing feature recognition. In this, fully connected VGG16 network (FC-VGG16 Net) efficiently handles sequential data with a one-dimensional structure, while convolutional VGG16 network (Conv-VGG16 Net) effectively models spatial–temporal sequences by leveraging convolution operations, enhancing the comprehensive modeling of spatial–temporal dynamics within video clip data. The STCU module explores the complementary nature of spatial and temporal features through an alternating co-attention mechanism, optimizing feature fusion. The classification stage labels video clips as violent or non-violent based on the extracted features. Together, these modules optimize the model's ability to discern complex patterns and salient features across spatial and temporal dimensions in video data. Experimental results showcase enhanced performance metrics, including increased precision and recall rates of overall 99.6%, validating its effectiveness in accurately classifying violent and non-violent actions on four video datasets.