Inertial Trajectory Estimation Using Low-Cost Inertial Measurement Units and Edge Computing

Trajectory estimation is often used in applications such as rehabilitation assessment and indoor navigation. Although various sensors can be used to estimate trajectories, inertial measurement units (IMUs) are both easy to implement and can be used in both indoor and outdoor environments. This study proposes a system in which trajectory estimation models are integrated into an edge computing platform with a neural processing unit to provide estimations without transmission delays. The system is secure, is private, and has low power consumption. A motion dataset comprising walking and hand movement data was collected, and a model with convolutional and temporal layers was trained and tested. Various tests were used to identify a suitable model structure and parameters; Res2Net, a convolutional block attention module, and a temporal convolutional network were ultimately used for the proposed model. Experiments were conducted to investigate the model's accuracy and inference time. The model's accuracy was high, with an average root-mean-square error of 0.364 m, exceeding that of a previously proposed model. Moreover, the model's inference time was 0.234 s for 20 s of input IMU data, which was >20% faster than that of a comparable ResNet bidirectional long short-term memory model. The model's performance was also evaluated on various platforms to demonstrate that it could provide accurate real-time trajectory estimation on an edge computing platform.