Patient activity recognition using radar sensors and machine learning

Indoor human activity recognition is actively studied as part of creating various intelligent systems with applications in smart home and office, smart health, internet of things, etc. Intrusive devices such as video cameras or sensors attached to the human body are often used to realize human activity recognition. These solutions, however, lead to various privacy issues. On the other hand, radar sensors are privacy-preserving and provide a lot of information about the subject such as speed, distance, range, and angle. Moreover, radar sensors can sense through the walls. In this respect, we investigate the use of radar data to achieve patient activity recognition. In particular, human activity data are collected from both an indoor environment that replicates a hospital setting and a real-life hospital room using two high dimensional radar sensors. The data are further fed to various supervised Machine Learning (ML) classification approaches. We investigate the robustness and generalization capabilities of the ML approaches with respect to people's age, radar sensor position, mobility aids and environments. The results show promising levels of accuracy. The Convolutional Neural Network (CNN) using Micro-Doppler (MD) maps are more effective for generalizing across different environments and radar positions with 62% and 73% accuracy, respectively. The CNNs using Range-Doppler (RD) maps are more efficient than using MD maps within the same environment in the case of distribution of age (87-95%), mobility aids (91-95%) and with different subjects (93-95%). A subset of the data set is made publicly available.Indoor human activity recognition is actively studied as part of creating various intelligent systems with applications in smart home and office, smart health, internet of things, etc. Intrusive devices such as video cameras or sensors attached to the human body are often used to realize human activity recognition. These solutions, however, lead to various privacy issues. On the other hand, radar sensors are privacy-preserving and provide a lot of information about the subject such as speed, distance, range, and angle. Moreover, radar sensors can sense through the walls. In this respect, we investigate the use of radar data to achieve patient activity recognition. In particular, human activity data are collected from both an indoor environment that replicates a hospital setting and a real-life hospital room using two high dimensional radar sensors. The data are further fed to various supervised Machine Learning (ML) classification approaches. We investigate the robustness and generalization capabilities of the ML approaches with respect to people's age, radar sensor position, mobility aids and environments. The results show promising levels of accuracy. The Convolutional Neural Network (CNN) using Micro-Doppler (MD) maps are more effective for generalizing across different environments and radar positions with 62% and 73% accuracy, respectively. The CNNs using Range-Doppler (RD) maps are more efficient than using MD maps within the same environment in the case of distribution of age (87-95%), mobility aids (91-95%) and with different subjects (93-95%). A subset of the data set is made publicly available.A