Scholarlink

<p>In today’s society, fighting behavior detection technology plays a vital role in preventing violent incidents and ensuring public safety. By combining surveillance cameras with cutting-edge object detection algorithms, real-time monitoring of crowd dynamics becomes achievable, enabling early intervention in potential conflicts. Based on the YOLO11 framework, SNEIF-YOLO11 introduces a lightweight ShuffleNetV2 network to replace the conventional backbone, significantly improving feature extraction speed while minimizing computational demands. The model incorporates an Efficient Local Attention module into its neck architecture, utilizing strip pooling and 1D convolutions to model long-range spatial relationships and address challenges such as occlusions and multi-scale feature variations. To optimize learning between simple and complex samples, the algorithm adopts Inner-FocalerIoU as its regression loss function, merging the boundary precision of Inner-IoU with the sample-adaptive weighting mechanism of Focaler-IoU. Experimental evaluations reveal that SNEIF-YOLO11 attains 98.4% mAP@0.5 and 79.3% mAP@50-95 on specialized violence detection datasets, while simultaneously reducing computational complexity (FLOPs) and model parameters. This advancement demonstrates strong potential for real-time identification of fighting behaviors in high-security surveillance environments. Dataset is available at: <a href="https://github.com/ZJY-HSBL/Fighting-Behavior">https://github.com/ZJY-HSBL/Fighting-Behavior</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

SNEIF-YOLO11: A Lightweight real-time framework for fighting behavior detection

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Wenshun Sheng,
  • Jiayu Zhao,
  • Qi Chen

摘要

In today’s society, fighting behavior detection technology plays a vital role in preventing violent incidents and ensuring public safety. By combining surveillance cameras with cutting-edge object detection algorithms, real-time monitoring of crowd dynamics becomes achievable, enabling early intervention in potential conflicts. Based on the YOLO11 framework, SNEIF-YOLO11 introduces a lightweight ShuffleNetV2 network to replace the conventional backbone, significantly improving feature extraction speed while minimizing computational demands. The model incorporates an Efficient Local Attention module into its neck architecture, utilizing strip pooling and 1D convolutions to model long-range spatial relationships and address challenges such as occlusions and multi-scale feature variations. To optimize learning between simple and complex samples, the algorithm adopts Inner-FocalerIoU as its regression loss function, merging the boundary precision of Inner-IoU with the sample-adaptive weighting mechanism of Focaler-IoU. Experimental evaluations reveal that SNEIF-YOLO11 attains 98.4% mAP@0.5 and 79.3% mAP@50-95 on specialized violence detection datasets, while simultaneously reducing computational complexity (FLOPs) and model parameters. This advancement demonstrates strong potential for real-time identification of fighting behaviors in high-security surveillance environments. Dataset is available at: https://github.com/ZJY-HSBL/Fighting-Behavior.