摘要
针对现有跌倒检测算法计算成本高、实时性差、难以在边缘设备上部署的问题,提出了一种基于YOLOv8n的轻量化跌倒检测模型WPD-YOLO。首先,在YOLOv8n中引入了Wise-IoU作为新的回归损失函数,以降低模型中低质量样本的负面影响,提升模型的收敛速度;其次,采用LAMP剪枝方法对改进后的模型进行压缩,从而有效降低模型的参数量和计算量;最后,通过通道知识蒸馏方法微调模型,在不额外增加参数量的情况下提升了模型的检测精度。实验结果表明,在公开的Fall-Detection数据集上,与YOLOv8n相比,WPD-YOLO的检测参数量减少了74.42%,计算量减少了50%,处理速度提升了56帧/s,同时平均准确率提升了0.4%,达到87.8%。
To address the issues of high computational cost,poor real-time performance,and difficulty in deployment on edge devices in existing fall detection algorithms,a lightweight fall detection model named WPD-YOLO based on YOLOv8n is proposed.First,Wise-IoU is introduced as a new regression loss function in YOLOv8n to reduce the negative impact of low-quality samples in the model and improve its convergence speed.Next,the LAMP pruning method is employed to compress the improved model,effectively reducing both the parameter count and computational load.Finally,the model is fine-tuned by channel knowledge distillation method,and the detection accuracy of the model is improved without additional parameters.Experimental results show that on the public Fall-Detection dataset,compared with YOLOv8n,the parameter count of WPD-YOLO is reduced by 74.42%,the computational complexity is reduced by 50%,the speed of processing is increased by 56 frames per second,while the mean average precision is increased by 0.4%,reaching 87.8%.
作者
华浩军
NAJI Alhusaini
HUA Haojun;NAJI Alhusaini(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232000,China;School of Artificial Intelligence,Chuzhou University,Chuzhou Anhui 239000,China)
出处
《盐城工学院学报(自然科学版)》
2025年第3期60-66,共7页
Journal of Yancheng Institute of Technology:Natural Science Edition
基金
2022年度安徽省高等学校科研计划项目(自然科学类重点项目)(2022AH051102)。
