欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (5): 791-802.DOI: 10.11996/JG.j.2095-302X.2022050791

• 图像处理与计算机视觉 • 上一篇    下一篇

基于优化 YOLOv5s 的跌倒人物目标检测方法

  

  1. 长安大学信息工程学院，陕西 西安 710064
  • 出版日期:2022-10-31 发布日期:2022-10-28
  • 基金资助:
    国家自然科学基金项目(51678061) 

An object detection method of falling person based on optimized YOLOv5s 

  1. School of Information Engineering, Chang’an University, Xi’an Shaanxi 710064, China
  • Online:2022-10-31 Published:2022-10-28
  • Supported by:
    National Natural Science Foundation of China (51678061) 

PDF (PC)

322

摘要：

针对目标检测模型在人物跌倒时易漏检、鲁棒性和泛化能力差等问题，提出一种基于改进 YOLOv5s 的跌倒人物目标检测方法 YOLOv5s-FPD。首先，对 Le2i 跌倒数据集使用多种方式扩充后用于模型 训练，增强模型鲁棒性和泛化能力；其次，使用 MobileNetV3 作为主干网络来进行特征提取，协调并平衡模型 的轻量化和准确性关系；然后，利用 BiFPN 改善模型多尺度特征融合能力，提高了融合速度和效率，并使用 CBAM 轻量级注意力机制实现注意力对通道和空间的双重关注，增强了注意力机制对模型准确性地提升效果； 最后，引入 Focal Loss 损失评价从而更注重挖掘困难样本特征，改善正负样本失衡的问题。实验结果表明，在 Le2i 跌倒数据集上 YOLOv5s-FPD 模型比原 YOLOv5s 模型，在精确度、F1 分数、检测速度分别提高了 2.91%， 0.03 和 8.7 FPS，验证了该方法的有效性。

关键词: 目标检测, YOLOv5s, MobileNetV3, 轻量级注意力, 多尺度特征融合, 焦点损失函数

Abstract:

To address the problems of easy missing, poor robustness and generalization ability when object detection model is detecting a person falling down, a new detection method YOLOv5s-FPD was proposed based on the improved YOLOv5s. Firstly, the Le2i fall detection data set was expanded in various ways for model training to enhance model robustness and generalization ability. Secondly, MobileNetV3 was employed as the backbone network for feature extraction, which could coordinate and balance the relationship between lightness and accuracy of the model. Furthermore, BiFPN (bi-directional feature pyramid network) was utilized to boost model multi-scale feature fusion ability, thereby improving the efficiency and speed of fusion. Meanwhile, the CBAM (convolutional block attention module) lightweight attention mechanism was adopted to realize double focus attention to channel and space, enhancing the effect of attention mechanism on model accuracy. Finally, Focal Loss evaluation was used to pay more attention to hard example mining and alleviate the samples imbalance problem. The experimental results show that the precision, F1 score, and detection speed of YOLOv5s-FPD model were improved by 2.91%, 0.03, and 8.7 FPS, respectively, compared with the original YOLOv5s model on Le2i fall detection dataset, which verified the effectiveness of the proposed method. 

Key words:  , object detection, YOLOv5s, MobileNetV3, lightweight attention, multi-scale feature fusion, focal loss function

中图分类号: