Integrated Utilization of IMU-based Human Activity Recognition Datasets Across Varied Configurations using MIG HAR Dataset

Abstract

Deep-learning-based IMU-based (Inertial measurement unit) Human Activity Recognition (HAR) has a problem of lack of large datasets. This problem would be solved if integrated utilization of small labeled datasets was established. However, this solution is not easy because there are feature-space differences among datasets depending on the IMU installation locations, recording environment, characteristics of used IMU, etc. To solve this problem, we adjusted the differences between datasets to enable their integrated utilization using the MIG HAR Dataset, consisting of synchronized data from 396 IMUs deployed throughout the whole body. Our approach adjusts sensor characteristic differences among datasets and among sensor positions by a transform method that is trained to adjust the characteristic differences among the MIG HAR’s sensors, selected by similarity between MIG HAR’s sensors and each sensor in each dataset. The results of an evaluation showed a macro-F1 score of 0.7915 when the MIG HAR, PAMAP2, SHO USC-HAD, and MotionSense datasets were used as training data, and the RealWorld dataset was used as test data. This score was also approximately 0.0321 points higher than the baseline macro-F1 score of 0.7594 obtained by leave-one-subject-out cross-validation of RealWorld.