| 摘 要: 随着经济社会的发展，人员跌倒尤其对于老年人以及危险工作者来说，已经成为一个巨大的隐患。针对现有人员跌倒检测算法存在模型参数量大、计算复杂度高以及轻量化与检测精度难以兼顾的问题，本文以YOLOv5s为基线模型，提出一种轻量化跌倒检测算法GCS-YOLOv5s。首先，在Backbone深层网络及Neck部分引入C3Ghost模块，通过Ghost机制减少冗余特征生成，在降低模型参数量与计算量的同时保持特征表达能力；其次，在检测头部分引入GSConv结构，以组卷积结合通道重排的方式增强轻量化条件下的特征交互能力；最后，采用SimSPPF替代原SPPF模块，通过串行池化结构扩大感受野并降低多尺度池化带来的计算开销。实验结果表明，相较于原始YOLOv5s模型，改进后的模型在mAP指标上提升了1.4个百分点，参数量减少2.5M，计算量降低4.4GFLOPs。 |
| 关键词: 人员跌倒 目标检测 深度学习 标准卷积 轻量级 多尺度池化 特征融合 |
|
中图分类号:
文献标识码:
|
|
| Lightweight Personnel Fall Detection Network Optimization |
|
lizhijin, wuhao
|
School of Artificial Intelligence, Nanjing University of Information Science and Technology
|
| Abstract: With the development of the economy and society, falls, especially among the elderly and workers in hazardous jobs, have become a significant hidden danger. Addressing the problems of existing fall detection algorithms, such as large model parameters, high computational complexity, and difficulty in balancing lightweight design with detection accuracy, this paper proposes a lightweight fall detection algorithm, GCS-YOLOv5s, based on the YOLOv5s model. Firstly, the C3Ghost module is introduced into the backbone deep network and neck part, reducing redundant feature generation through the Ghost mechanism, thereby decreasing model parameters and computation while maintaining feature representation capabilities. Secondly, the GSConv structure is introduced in the detection head, enhancing feature interaction under lightweight conditions through group convolution combined with channel shuffle. Finally, SimSPPF is used to replace the original SPPF module, expanding the receptive field and reducing the computational overhead caused by multi-scale pooling through a serial pooling structure. Experimental results show that, compared with the original YOLOv5s model, the improved model increases the mAP by 1.4 percentage points, reduces parameters by 2.5M, and decreases computation by 4.4 GFLOPs. |
| Keywords: Personnel falls Object detection Deep learning Standard convolution Lightweight Multi-scale pooling Feature fusion |
