基于改进Yolov8的摔倒行为检测算法
Fall behavior detection based on improved Yolov8 algorithm
白亮 1丁学文 2申明坤 1王树云 1常黎玫1
作者信息
- 1. 天津职业技术师范大学电子工程学院,天津 300222
- 2. 天津职业技术师范大学电子工程学院,天津 300222;天津云智通科技有限公司,天津 300350
- 折叠
摘要
为防止发生摔倒行为后得不到及时救助而产生危害生命的风险,提出一种改进Yolov8 的摔倒行为检测算法Yolov8-DCN-WIoU-C3P.通过使用可随目标形态进行自适应采样的可变卷积(Deformable Convolutional Networks,DCN),来提高C2f模块(CSP Bottleneck with 2 convolutions,C2f)对不规则摔倒目标的特征提取能力;引入针对交叉熵的单调聚焦机制的损失函数WIoU,来提高算法对低质量摔倒数据的泛化能力;采用减少冗余计算和内存访问的部分卷积Pconv,来解决像素损坏导致计算量上涨的问题,进而使算法识别摔倒行为的性能得到提升.实验结果表明:本文提出的改进算法能有效识别出摔倒行为,对比原始算法(Yolov8算法)在准确率以及平均精度上分别提高了1.6%和2.1%.
Abstract
To prevent the risk of potential life-threatening situations due to delayed assistance following a fall,this pa-per proposes the YOLOV8-DCN-WIou-C3P,an improved Yolov8 fall detection algorithm.This algorithm enhances the feature extraction capability of the C2f module(CSP Bottleneck with 2 convolutions,C2f)for irregular falling objects by utilizing deformable convolutional networks(DCN)that allow adaptive sampling based on the target's shape.The loss function WIoU based on the monotony focusing mechanism of cross entropy is used to improve the generalization capa-bility of the algorithm to low-quality fall data.Partial convolutions(Pconv)are used to reduce redundant computations and memory access,addressing the issue of increased computational load due to pixel corruption,thereby enhancing the performance of fall detection.The results show that the improved algorithm can effectively identify falls,achieving 1.6%increase in accuracy and 2.1%increase in average precision compared to the original Yolov8 algorithm.
关键词
摔倒行为/Yolov8n/可变卷积/WIoU/卷积检测算法Key words
fall/Yolov8n/deformable convolution/WIoU/convolutional detection algorithm引用本文复制引用
基金项目
天津市科技计划(20YDTPJC01110)
出版年
2024
NETL
NSTL
万方数据