向勉, 易本顺, 周丙涛, 谭建军, 朱黎. 利用惯性传感器与多模态网络解析跑步参数[J]. 武汉大学学报 ( 信息科学版), 2024, 49(7): 1079-1087. DOI: 10.13203/j.whugis20220229
引用本文: 向勉, 易本顺, 周丙涛, 谭建军, 朱黎. 利用惯性传感器与多模态网络解析跑步参数[J]. 武汉大学学报 ( 信息科学版), 2024, 49(7): 1079-1087. DOI: 10.13203/j.whugis20220229
XIANG Mian, YI Benshun, ZHOU Bingtao, TAN Jianjun, ZHU Li. Analysis of Running Parameters Using IMU and Multi-modal Network[J]. Geomatics and Information Science of Wuhan University, 2024, 49(7): 1079-1087. DOI: 10.13203/j.whugis20220229
Citation: XIANG Mian, YI Benshun, ZHOU Bingtao, TAN Jianjun, ZHU Li. Analysis of Running Parameters Using IMU and Multi-modal Network[J]. Geomatics and Information Science of Wuhan University, 2024, 49(7): 1079-1087. DOI: 10.13203/j.whugis20220229

利用惯性传感器与多模态网络解析跑步参数

Analysis of Running Parameters Using IMU and Multi-modal Network

  • 摘要: 实时检测跑步时的速度与步幅在避免运动者受伤、提升运动效率上有着重要意义。提出了一种利用惯性测量单元(inertial measurement unit,IMU)来检测这两个指标的方法。首先,招募了10名志愿者,并将3个IMU模块安置在足部、小腿、大腿处,采集了5 137个步态周期的数据;然后,利用主成分分析法分析数据,结合皮尔逊相关系数探讨了速度步幅与传感器位置、物理参数之间的关系;提出了一种多模态架构的长短期记忆特征提取网络(multi-modal-attention-long short-term memory,M-Att-LSTM),利用两个引入注意力机制的长短期记忆网络(att-long short-term memory,Att-LSTM)对加速度和角度变化分别做特征提取,最后进行回归拟合。实验结果表明,M-Att-LSTM在速度上误差为0.058 m/s,标准偏差为0.013 m/s,而在步幅上误差为0.023 m,标准偏差为0.022 m,两项指标都优于单纯的Att-LSTM。

     

    Abstract:
    Objectives Real-time measurement of running speed and stride length is of great significance in avoiding injury and improving exercise efficiency.
    Methods We propose a method using inertial measurement unit (IMU) to detect these two indicators. First, 3 IMU are placed on the foot, calf and thigh of the 10 runners which we recruited, and 5 137 data of gait cycles are collected. Second, principal component analysis is used to analyze the data, and Pearson correlation coefficient is used to discuss the relationship between the detection indicators of running and the sensor position and physical parameters. Then a multi-modal attention-long short-term memory (M-Att-LSTM ) is proposed for feature extraction, two long short-term memory (LSTM) modules with attention mechanism are used to extract features of acceleration and angle, and regression fitting is carried out.
    Results The experiment result shows that M-Att-LSTM has errors of 0.058 m/s in speed and 0.023 m in stride, the standard deviation is 0.013 m/s and 0.022 m, respectively.Both indicators are better than pure Att-LSTM.
    Conclusions The studies show that multi-modal network can improve network processing capabilities, compared with relevant researches in recent years, our study has obvious advantages in error control.

     

/

返回文章
返回