Lightweight Neural-Network-Based Trajectory Estimation for Low-Cost Inertial Measurement Units

Generally, inertial measurement unit can measure the acceleration and angular velocity of an object in three-dimensional space, and use this to calculate the object's attitude and movement trajectory. In particular, motion trajectories can be applied to rehabilitation assessment, but they have problems with large computational models and slow calculation speeds. In order to solve the above problems, a lightweight neural network model for trajectory estimation was proposed. Res2Net and temporal convolution network (TCN) were used for extract spatial and temporal features from inertial data. Tests using various datasets, window sizes, and batch sizes were held. The results show that a window size of 100 and a batch size of 16 is most suitable for our computing system. The proposed model reduces inference time by 50.3% compared to previous best work proposed by Lin et al. in 2022 while maintaining adequate accuracy. The model has low model size while could be easily implemented in smartphone and edge computing platforms.