mmWave Radar Point Cloud Based Pose Estimation with Residual Blocks for Rehabilitation Exercise

Rehabilitation exercise is essential for recovery from injuries, surgeries, or physical impairments. Radar-based pose estimation offers detailed insights into skeletal structures and joint movements during rehabilitation. Traditional methods project radar point cloud data into image-like formats for feature extraction via Convolutional Neural Network (CNN). However, this can result in spatial information loss during downsampling. To address this, we integrate residual blocks with CNN. Our approach begins with CNN extracting local features from projected point clouds, followed by skip connections after each convolutional layer to ensure continuous information flow. This method improves learning from radar data, aiding in capturing both global and local pose details. Experiments on an open-source mm-wave radar dataset confirm our method’s effectiveness, with an average localization error of 6.37 cm.