| 摘 要: 为了降低跌倒所致并发症或伤亡的风险,使用深度学习方法中ResNet18(Residual Network 18)经典卷积结构对传感器时间序列数据进行处理,采用10折交叉验证探究滑动时间窗口和不同输入特征对模型跌倒检测性能的影响。实验结果表明,为MobiAct2.0、SisFall、Cogent Labs数据集选择适宜的窗口时,跌倒检测准确率均达到92%以上。此外,数据集获取差异对特定模型的分类性能具有显著影响。本研究证实,基于ResNet18的网络架构在传感器数据跌倒检测中具有一定的应用潜力,但对复杂跌倒场景的识别能力需探索更优的数据融合和特征处理方法以提升模型的检测性能。 |
| 关键词: ResNet18;传感器时序数据;公共数据集;跌倒检测;输入特征 |
|
中图分类号: TP391
文献标识码: A
|
|
| Application of ResNet18 in Fall Detection on Sensor-Based Public Datasets |
|
WAN Pengbo, SHI Yujiao, ZHAO Yizhu, LI Xueqing
|
(College of Art and Design, Shaanxi University of Science & Technology, X'i an 710021, China)
Wanpengbo@yeah.net; Syj13892681635@163.com; 1394721662@qq.com; 2929696278@qq.com
|
| Abstract: In order to reduce complications or casualties caused by falls, this paper proposes to process sensor time series data with the ResNet18 classical convolutional structure in a deep learning approach, and then use 10-fold cross-validation to explore the effect of sliding time windows and different input features on the model performance of fall detection. Experimental results demonstrate that when appropriate windows are selected for the MobiAct2. 0, SisFall, and Cogent Labs datasets, fall detection accuracy exceeds 92% . Moreover, the differences in dataset acquisition significantly affect the classification performance of specific models. This research verifies that the ResNet18-based network architecture holds considerable application potential in fall detection using sensor data; however, the ability to recognize complex fall scenarios necessitates exploring superior data fusion and feature processing methods to enhance the model's detection performance. |
| Keywords: ResNet18; sensor time series data; public datasets; fall detection; input features |
