The Impact of Environmental Factors on mm-Wave Radar Point-Clouds for Human Activity Recognition

Recently, the millimeter wave (mmWave) radar sensing has attracted significant attention due to the physical characteristic of mmWave signals and the large 5G frequency bands. Transforming the mmWave signals into point clouds via physics enables many new applications such as human activity recognition. However, learning the human activity from the mmWave point-clouds are susceptible to many environmental/dynamic factors, such as the spatial diversity, facing orientation, and the physical stature of users, which can severely degrade the performance of radar-based human activity recognition systems. By developing a dataset based on the TI hardware platform, this paper builds a baseline recognition system using convolutional neural networks [1], investigates the properties of mmWave point-clouds, and reports the recognition accuracy for six human activities under different experimental scenarios including the distinct testing locations, different orientations and physical stature of users.