Radar-Based Human Skeleton Estimation with CNN-LSTM Network Trained with Limited Data

This paper presents a novel framework for human pose estimation using millimeter-wave (mmWave) radar technology, focusing on personalized healthcare applications. The proposed framework utilizes range-azimuth, range-elevation, and range-Doppler maps as inputs to a convolutional neural network (CNN) with a long short-term memory (LSTM) architecture to capture temporal dependencies and achieve improved skeleton estimation. Furthermore, this paper addresses the limitations of current radar-based skeleton estimation techniques, such as inconsistent kinematics and reliance on sparse radar point clouds. Skeleton estimation accuracy attained using diversified simulations is compared with that achieved real RF data, as validated using gold standard Vicon motion capture (MoCap) measurements as ground truth. The results highlight the potential of mmWave radar-based human skeleton estimation for advancing personalized healthcare and improving gait analysis and fall risk assessment.