Part of MONSTER: https://arxiv.org/abs/2502.15122.

USCActivity consists of data collected from a Motion-Node device, which includes six readings from a body-worn 3-axis accelerometer and gyroscope sensor [1]. The dataset contains samples from 14 male and female subjects with equal distribution (7 each) and specific physical characteristics and ages. The sensor data is sampled at a rate of 100 Hz, and each time-step in the dataset is labeled with one of 12 activity classes.

The USCActivity dataset presents a challenge in learning feature representation and segmentation due to the placement of the sensors and the variability in activity classes. The data is collected from a single accelerometer and gyroscope reading obtained from a motion node attached to the subject's right hip. Therefore, this reading does not contribute significantly to the feature space transformation. Additionally, the activity classes involve various orientations, such as walking forward, left, or right, and even using the elevator up or down, which cannot be captured solely through accelerometer and gyroscope readings. Similar to other activity recognition datasets, we use subject-based cross-validation.

[1] Mi Zhang and Alexander A Sawchuk. (2012). USC-HAD: A daily activity dataset for ubiquitous activity recognition using wearable sensors. In Conference on Ubiquitous Computing, pages 1036–1043.