DNN Model Comparison for Sensor Location Robustness

Many IMU(inertial measurement unit) based human activity recognition (HAR) models have been developed. However, these models need to be more robust to misalignment of the IMU mounting position. Therefore, we compared and verified the robustness of existing IMU-based HAR models in terms of classification performance when the mounting position differs between training and estimation to clarify the characteristics of each method. We used the Meshed IMU Garment HAR Dataset (MIGHAR Dataset) for the evaluation. We compared the decline in recognition performance due to the distance between IMUs using five different models: DeepConvLSTM, ICGNet, iSPLInception, mobileHART, and rTsfNet. As a result, all models showed a decrease in recognition accuracy. However, the rTsfNet was the most robust, maintaining 55.04%PSIMU (Performance Percentage compared to training/evaluation using the Same IMU) even at the maximum Manhattan distance and showing a Macro F1-score (mf1) of 0.4384\pm0.0672. Second place went to the mobileHART, with 37.56%PSIMU and 0.2801\pm0.0526 mf1 at the maximum Manhattan distance. The DeepConvLSTM and the ICGNet were about the same, with 28.34%PSIMU and 29.16%PSIMU the maximum Manhattan distance and 0.1828\pm0.0675 and 0.1900\pm0.0423 mf1, and the iSPLInception achieved 22.75%PSIMU and 0.1712\pm0.0518 mf1. The conclusion was that dynamic 3D rotation of sensor data built-in rTsfNet ensures robustness against IMU mounting position errors.