Millimeter-wave (mmWave) radar is emerging as a privacy-preserving technology for in-home monitoring, particularly for older adults who prefer aging in place. This study addresses the need for accurate human action recognition (HAR) in real-world settings by introducing a dataset of natural activities collected from 28 homes. The dataset captures sit-down and stand-up actions--key indicators of mobility--using the Chirp Smart Home Sensor, which combines sparse 3D point cloud data from an mmWave radar with low-resolution thermal imaging for annotation.
The dataset comprises over 900 annotated actions across diverse residential environments, supplemented with augmentation techniques to mitigate class imbalance. A baseline convolutional neural network (CNN) model is evaluated using Doppler-time (DT) and positional-time data (XT, YT, ZT) features. Results reveal significant challenges, including high variability between training and testing environments, low precision due to diverse non-action classes, and limited spatial coverage of action locations.
PDF