桂林理工大学学报2025，Vol.45Issue(2)：251-259,9.DOI:10.3969/j.issn.1674-9057.2025.02.014
基于梅尔谱特征和改进ResNet网络的室内跌倒检测方法
Indoor fall detection method based on Mel-spectrum feature and improved ResNet
摘要
Abstract
To overcome the limitations of existing methods for identifying falls among the elderly,a novel tech-nique is proposed that utilizes sound signals for fall detection.In the acoustic feature extraction stage,Mel-fre-quency cepstral coefficients(MFCC)are supplemented with temporal information.The Mel-frequency cepstral coefficients along with their first and second-order differential coefficients are transformed into a three-dimen-sional feature similar to an image,which is then classified using convolutional neural networks,thereby enhan-cing the robustness of indoor fall detection against noise.The network architecture is further improved through the use of the SimAM attention mechanism,feature pyramid network(FPN),and dynamic receptive convolution(DRConv).Experimental results show that this proposed method outperforms traditional identification methods on different datasets.The refined network model achieves precision,recall,and F1-Score of 98.43％,98.21％,and 98.32％,respectively,on the A3FALL dataset.Furthermore,for human fall sound events,the F1-Score is 96.45％,which is superior to that of other conventional network models.关键词
跌倒检测/SimAM/卷积神经网络/特征金字塔/动态区域感知卷积/梅尔频率倒谱系数(MFCC)Key words
fall detection/SimAM/convolutional neural network/FPN/DRConv/MFCC分类
信息技术与安全科学引用本文复制引用
杨松铭,王玫..基于梅尔谱特征和改进ResNet网络的室内跌倒检测方法[J].桂林理工大学学报,2025,45(2):251-259,9.基金项目
国家自然科学基金项目(62071135 （）
61961010) （）
广西自然科学基金项目(2019GXNSFBA245103) （2019GXNSFBA245103）
