人体动作识别中的惯性传感器位置校正算法

Position correction algorithm of inertial sensors in human action recognition

中文关键词  传感器系统  位置校正  旋转矩阵  动作识别  加权BP神经网络

英文关键词  sensor system  position correction  rotation matrix  action recognition  weighted BP neural network

中文摘要

      近年来, 惯性传感器在人体动作识别中的应用受到了广泛关注, 但用户在重新穿戴惯性传感器时, 不能保 证每次的固定位置完全一致, 这会影响识别精度. 针对此问题, 本文提出一种利用旋转矩阵实现惯性传感器位置校 正的人体动作识别方法. 首先, 将惯性传感器按照不同的位置固定在手腕处采集动作数据. 然后, 根据矩阵旋转变换 原理, 通过标准固定位置与其他固定位置的基准数据求取旋转矩阵. 最后, 对动作数据提取时频域特征, 并构造加 权BP神经网络模型以验证校正方法的有效性. 同时还讨论了不同的数据融合方法对动作识别的影响. 结果表明, 校 正后的动作数据的识别准确率分别为84.25%, 85.94%, 相比校正前提高了66.16%, 54.35%, 说明该方法是有效的.

英文摘要

      In recent years, the application of inertial sensors in human action recognition has received extensive attention, but users cannot guarantee that the fixed positions are completely consistent each time they wear inertial sensors again, which can affect the recognition accuracy. Aiming at this problem, this paper proposes a human action recognition method using rotation matrix to realize the position correction of inertial sensors. Firstly, the inertial sensors are fixed at the wrists according to different positions to collect action data. Secondly, the rotation matrices are obtained from benchmark data of the standard fixed position and other fixed positions according to the principle of matrix rotation transformation. Finally, the time-frequency domain features of the action data are extracted, and the weighted back propagation (BP) neural network model is constructed to verify the effectiveness of the correction method. It also discusses the influence of different data fusion methods on action recognition. As a result, the action accuracies of the corrected test data are 84.25%, 85.94%, which are increased by 66.16%, 54.35% compared with those before correction. It shows that the method is effective.