基于DirtNet与惯性测量单元的人体姿态估计

Human Pose Estimation Based on DirtNet and Inertial Measurement Units

摘要: 仅使用少量的惯性测量单元(IMU, Inertial Measurement Unit)进行人体姿态估计是一种非侵入性且经济的人体姿态估计方法，该方法主要面临的挑战是从带有噪声的IMU信号中精确估计人体姿态。为此，对人体姿态估计问题提出了一种仅使用6个IMU精确估计人体姿态的方法。1) 提出了一种双重信息保留注意力Transformer网络(DirtNet, Dual information retention transformer Network)，它能够有效保留历史信息并通过注意整个序列的信息来获得更好的结果。2) 通过对加速度进行积分了获得了近似变化速度，并将其作为额外的输入通道以提高了人体姿态估计的精确度。3) 使用均匀滤波过滤和白噪声模拟的方法对合成的加速度进行了数据增强，以此来拟合真实的IMU数据并得到更好的训练结果。与之前的研究相比，改进后的方法有效提高了姿态估计的精确度。

Abstract: Using a small number of inertial measurement units (IMUs) for human pose estimation is a non-intrusive and cost-effective method. However, accurately estimating human pose from noisy IMU signals poses a significant challenge. To address this challenge, a method that utilizes only six IMUs for precise human pose estimation is proposed. 1) A dual information retention attention Transformer network, called DirtNet, is introduced. This network effectively preserves historical information and leverages attention over the entire sequence to achieve better results. 2) The approximate velocity is obtained by integrating the acceleration, and it is used as an additional input channel to improve the accuracy of human pose estimation. 3) A data augmentation technique is applied by filtering the synthesized acceleration using uniform filtering and simulating white noise. This approach helps to fit the real IMU data and achieve better training results. Compared to previous research, the improved method significantly enhances the accuracy of pose estimation. By combining the strengths of DirtNet, leveraging historical information, incorporating velocity as an input, and applying data augmentation techniques, this method provides more precise human pose estimation results.