Radar-Based Human Activity Recognition Using Adaptive Range Selection and Deep Neural Network

Radar-based human activity recognition (HAR) supports monitoring systems with high user convenience and low privacy risk. The time-Doppler (TD) domain, excelling in capturing subtle human motions, is commonly employed in radar-based HAR. However, the process of converting a raw signal into this domain dilutes activity features and accumulates noise due to range-axis summation. Therefore, we propose an adaptive range selection (ARS) and range-sequential HAR network (RS-HARNet) for HAR based on range-time-Doppler (RTD) map. The ARS estimates the optimal range bin by evaluating the cross-correlation between each RTD map and a TD map containing activity features. The RS-HARNet consists of an encoder that extracts features from each RTD map and a decoder that extracts sequential features from consecutive RTD maps. To validate the performance of the proposed method, we compared two RTD map input formats and conducted k-fold cross-validation on a publicly available radar-based HAR dataset. Experimental results demonstrate that the proposed method achieves high activity classification by effectively leveraging the activity feature.