Acquisition of Unlabeled Dataset for Human Activity Recognition

In the paper, we recorded the device’s IMU sensors in the setting of Human Activity Recognition task for the purpose of building the pretrained model by self-supervised learner. To use the data for self-supervised learning, we collected it without labels. Therefore, unlike typical human activity recognition datasets, no labels were assigned, and it is important to note this distinction. As a result, the cost of label collection was minimized, but the time required for data collection remained the same, even though the data was unlabeled. On the other hand, since the data was unlabeled, there was no need to go through the complex process of assigning labels or inputting labels through a smartphone, which meant that data collection could be carried out without any prior knowledge of human activity recognition. The dataset consists of data from three subjects, totaling 228 hours. We performed self-supervised learning on the model and evaluated its performance using other IMU datasets. While this data was originally collected for human activity recognition, we anticipate that the use of unlabeled data in self-supervised learning will become more common in the future. In such cases, this dataset should be suitable for tasks that use unlabeled IMU data.