基于轻量化网络的帕金森步态识别方法  

Parkinson’s gait recognition method based on lightweight network

作　　者：郭坛 时文雅 郇战[2] 刘洋[1] GUO Tan;SHI Wenya;HUAN Zhan;LIU Yang(College of Computer and Artificial Intelligence,Changzhou University,Changzhou 213000,China;College of Microelectronics and Control Engineering,Changzhou University,Changzhou 213000,China)

机构地区：[1]常州大学计算机与人工智能学院,江苏常州213000 [2]常州大学微电子与控制工程学院,江苏常州213000

出　　处：《传感器与微系统》2025年第4期143-147,共5页Transducer and Microsystem Technologies

基　　金：国家自然科学基金资助项目(12201075)。

摘　　要：为了提高帕金森步态的识别效率并保持高识别精度,提出了一种基于轻量化帕金森步态识别方法-多头量化时域卷积网络(MQ-TCN)。用TCN层替换深度可分离卷积中的逐通道卷积,并部署TTQ算法,减少模型的参数量和参数复杂度。其次,该研究还分析了帕金森步态数据的冗余性,在略微损失识别精度的前提下大幅降低了模型训练所需的存储空间,进一步提升了模型在轻量设备中的可部署能力。实验结果显示:改进的MQ-TCN平均识别精度达到94.9%,参数量仅为目前最小帕金森步态识别模型的5%,不但保持高效的识别精度,还大幅度降低了模型的参数量与参数复杂度,为后续帕金森步态识别工具在轻量设备上的部署提供了参考依据。In order to improve the recognition efficiency of Parkinson’s gait and maintain high recognition precision,a lightweight Parkinson’s gait recognition method based on multi-head quantized temporal convolution network(MQ-TCN)is proposed.The TCN layer is used to replace the channel-by-channel convolution in the depth separable convolution,and the trained ternary quantization(TTQ)algorithm is deployed to reduce the number of parameters and parameter complexity of the model.Secondly,this study also analyzes the redundancy of Parkinson’s gait data,which greatly reduces the storage space required for model training under the premise of slightly losing recognition precision,and further improves the deployable ability of the model in lightweight devices.The experimental results show that the average recognition precision of the improved MQ-TCN reaches 94.9%,and the quantities of parameters are only 5%of the current minimum Parkinson’s gait recognition model.It not only maintains efficient recognition precision,but also greatly reduces the quantities of parameters and parameter complexity of the model,which provides a basis of reference for the subsequent deployment of Parkinson’s gait recognition tools on lightweight devices.

关 键 词：异常步态识别 轻量化卷积 时域卷积网络 参数量化 模型压缩 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]