Human Activity Recognition Trained on Simulated Millimeter-Wave Radar Data With Domain Adaptation

Most approaches for human activity recognition (HAR) with millimeter-wave radar are based on machine learning algorithms trained with extensive data. This strategy, while prevalent, requires a considerable investment of time and work-force for data acquisition, highlighting the urgent need for alternative ways of generating training samples. This paper presents a simulation-augmentation of training data with domain adaptation. It first introduces an innovative simulation framework based on Frequency-Modulated Continuous-Wave (FMCW) radar, designed to generate simulated radar data depicting human activities. To address the systematic differences between simulated and real-world measured data, which significantly reduce their effectiveness as auxiliary data resources, this study proposes the novel domain adaptive generation-recognition network (DAGRN), developed to integrate the critical features of measured data into the simulated data domain. It improves the data efficiency and quality within the scope of HAR. The results show that the approach achieves up to a 99.0% recognition accuracy rate across various positional scenarios for entirely new subjects. Notably, the training of the recognition module solely utilized the generated simulated-to-measured adaptation data. Yet, identification accuracy surpasses methods that rely exclusively on simulation, measurement, or a hybrid of both data types.