Semi-supervised animal action segmentation: HuGaDB

The Human Gait Database (HuGaDB) is a publicly available action segmentation dataset that contains lower limb activities such as walking, running, and sitting for a total of 18 subjects (Chereshnev et al. 2017, AIST). HuGaDB measured movements from six inertial measurement units (IMUs) with a 60 Hz sampling frequency. The IMUs were placed on each individual’s right and left feet, shins, and thighs. Each inertial sensor tracks the acceleration data and gyroscope data on each of the three x,y,z axes. We consider eight behavioral categories: walking, running, going up, going down, sitting, sitting down, standing up, and standing. The original dataset contained four additional behaviors that we excluded from our analysis. Two of those behaviors, “bicycling” and “sitting in car,” were not present in the publicly available dataset. The IMU data for the classes “up by elevator” and “down by elevator” did not seem to contain measurable variability.The dataset contains 517k frames, all of which are labeled.