Bio-Inspired Human Action Recognition With a Micro-Doppler Sonar System.

This paper explores computational methods to address the problem of doing inference from data in multiple modalities, where there exists a large amount of low dimensional data complementary to a much smaller set of high dimensional data. In this instance the low dimensional time-series data are active acoustics from a bio-inspired micro-Doppler sonar sensor system that include no or very limited spatial information, and the high dimensional data are RGB-depth data from a 3-D point cloud sensor. The task is human action recognition from the active acoustic data. To accomplish this, statistical models, trained simultaneously on both the micro-Doppler modulations induced by human actions and symbolic representations of skeletal poses, derived from the 3-D point cloud data, are developed. This simultaneous training enables the model to learn relations between the rich temporal structure of the micro-Doppler modulations and the high-dimensional pose sequences of human action. During runtime, the model relies purely on the active acoustic sonar data to infer the human action. Our approach is applicable to other sensing modalities, such as the millimeter wave electromagnetic radar devices.