基于加速度计数据的儿童身体活动类型识别模型构建与应用

Construction andApplication of Identification Model for Children's PhysicalActivity Types Based on Accelerometer Data

目的通过构建高准确率、高时间分辨率的儿童身体活动类型识别模型,展示活动类型、活动时间、活动强度深度融合的身体活动评价方式,为实现身体活动评价的多视角、可视化、可追踪提供思路。方法基于10种儿童身体活动的公开数据集,使用Python 3.8、Tensorflow 2.4构建识别身体活动类型的残差卷积神经网络模型并进行评估;将模型应用于自主设计的身体活动评价程序,基于活动类型识别、活动强度计算输出身体活动案例评价结果。结果残差卷积神经网络模型在区分跳绳与走上楼梯、静止、快跑、慢跑、走下楼梯、快走、慢走、坐下去、站起来等9种儿童身体活动类型时准确率达到99.3%,模型识别活动案例的准确率也达到99.1%,模型时间分辨率为2.8 s。结论模型的高准确率、高时间分辨率为儿童身体活动类型识别在身体活动评价中发挥重要作用奠定了坚实基础,可以促使身体活动评价更加全面、直观、精准。

Objective To construct a high-accuracy and high-time resolution model for the identification of children's physical activity type,in order to create conditions for exerting the role of the type of physical activity in physical activity evaluation.Meanwhile,the model is applied to the self-designed program for the evaluation of children's physical activity,to be used for displaying the evaluation method of physical activity in which activity type,time and intensity were integrated.Methods Open data sets of ten kinds of children's physical activity were used.Residual convolution neural network model was constructed and evaluated with Python 3.8 and Tensorflow 2.4.The model was applied to the self-designed program for physical activity evaluation for the output of the results of physical activity evaluation based on the activity type identification,activity count and intensity grade division results of the case.Results The accuracy of residual convolution neural network model in distinguishing 9 types of children's physical activities(jump rope and stair up,etc.)is 99.3%,the accuracy in physical activity's case is 99.1%,and the time resolution is about 2.8 seconds.Conclusion The high-accuracy and high-time resolution model plays a vital role in the identification of children's physical activity type of physical activity evaluation,which may help a more comprehensive,straight and accurate evaluation method of physical activity.