Radar Based Human Activity Recognition: from Classification to Detection

FMCW radar is emerging as a key technology for contactless Human Activity Recognition (HAR) in medical settings. While effective in controlled environments, FMCW radar-based HAR faces challenges in realistic scenarios, particularly with continuous data streams. Tradi- tional methods often rely on segmentation, which can be cumbersome, or fail to handle transitions between activities, leading to errors. This study introduces a novel approach that enhances activity detection by refining classifier outputs with simple probabilistic reasoning that takes into account activity transition probabilities. The proposed method im- proves the HAR detection F1 score with 6% compared to the baseline and even slightly improves the classification accuracy which does not take into historical information about previous activities.