The feasibility of recognizing different human–human interactions based on the time-range dimension is investigated. Measured data of six kinds of human–human interactions including hugging, kicking, pointing, punching, pushing, and shaking hands were collected by ultra-wideband radar (UWB). By observing the variation of the time-range dimension matrix, the time-varying UWB signatures are characterized in a time window. Four features are extracted from a time-range dimension matrix by analyzing the characteristics of time-varying UWB signatures. K-nearest neighbor is used to recognize six kinds of human–human interactions based on four measurement features, and the recognition accuracy is found to be up to 99%. In addition, the results obtained by the proposed recognition algorithm are compared with those obtained by other feature extraction algorithms, which further demonstrates the superiority and generalization ability of the algorithm. In addition, the recognition accuracy of the proposed algorithm is higher than some deep learning algorithms, including AlexNet, VGGNet, ResNet, and DenseNet in the case of small samples.