Scholarlink

Human object detection in video has significant value in the fields of security, competitive sports, sports training, and other fields. Although the existing technology can realize the high-precision analysis of human motion video, it seriously relies on the high resolution and high frame rate video recording of the camera. It is prone to error detection affecting the accuracy of motion analysis. To improve the target detection capability of human motion video at low resolution and low frame rate, this paper proposes a novel model YOLOv8n-GFE with high detection accuracy and low hardware requirements. The model introduces the Global Attention Mechanism (General Attention Module) to improve global perception and comprehensive extraction ability of features, optimizes the computational efficiency through deep separable convolution and merging convolution operations, and increases the accuracy and its ability to generalize complex environments through Focal and Wise-IoU loss functions. From the experiments, we see that YOLOv8n-GFE improves the accuracy by 3.25% over the original YOLOv8n model and by 2.048% over the COCO data set, confirming its effectiveness in video motion target detection, and also has significant advantages over other YOLO algorithms.

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

Moving Video Object Detection Method Integrating Attention Mechanism and Convolution Operation Optimization Strategy

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jian Yang,
  • Chunfang Zhou,
  • Guofu Deng,
  • Guo Li

摘要

Human object detection in video has significant value in the fields of security, competitive sports, sports training, and other fields. Although the existing technology can realize the high-precision analysis of human motion video, it seriously relies on the high resolution and high frame rate video recording of the camera. It is prone to error detection affecting the accuracy of motion analysis. To improve the target detection capability of human motion video at low resolution and low frame rate, this paper proposes a novel model YOLOv8n-GFE with high detection accuracy and low hardware requirements. The model introduces the Global Attention Mechanism (General Attention Module) to improve global perception and comprehensive extraction ability of features, optimizes the computational efficiency through deep separable convolution and merging convolution operations, and increases the accuracy and its ability to generalize complex environments through Focal and Wise-IoU loss functions. From the experiments, we see that YOLOv8n-GFE improves the accuracy by 3.25% over the original YOLOv8n model and by 2.048% over the COCO data set, confirming its effectiveness in video motion target detection, and also has significant advantages over other YOLO algorithms.