Integrated Framework for Human Activity Classification and Frame Reconstruction Using In-Pocket FMCW Radar

초록

We present an integrated framework for human activity classification and missing-frame reconstruction using an in-pocket FMCW radar sensor. We first employ a 2D convolutional neural network (CNN) + long short-term memory (LSTM) layer to classify five activities-sitting, standing, walking, running, and stair climbing-from range-Doppler maps, achieving 98.28% test accuracy under 5-fold validation. To address data loss from interference, we simulate frame drop rates of 30%, 50%, and 70% and benchmark three reconstruction strategies: set A (raw missing data), set B (preceding-frame copy), and set C (3D autoencoder). Across all activities, set B and set C show the improved average accuracy of 40~60 % relative to the base line, with set B edging set C by 1~2 %due to static movements. In the running case, set C outperforms set B by approximately 5% at 50 % drop and 15 % at 70 % drop, demonstrating the value of learned temporal features for dynamic action recovery. The results confirm the framework's robustness for portable radar-based activity monitoring. © 2025 IEEE.

키워드