DIHARNet: A Temporal-Doppler-Spatial domain Fusion Method for Direction Insensitive Human Activity Recognition based on MMWave Radar

Recently, millimeter-wave radar has been increasingly applied to Human Activity Recognition (HAR) due to its inherent privacy protection and resistance to low-light interference. However, methods relying solely on Time-Doppler Map (TDM) suffer from severe performance degradation as the subject's direction deviates from the radar's line-of-sight. To address this, we propose DIHARNet, a dual-stream framework designed for direction-insensitive recognition, which concurrently exploits the intrinsic micro-motion dynamics of TDM and the geometric robustness of Point Cloud Data (PCD). Specifically, we circumvent the prohibitive costs of dense volumetric learning by factorizing high-dimensional spatial data into orthogonal 2D projections. Furthermore, to rectify directional ambiguity, we introduce a geometry-guided recalibration mechanism that leverages direction-invariant spatial priors to dynamically calibrate Doppler features, ensuring discriminative representations across representative directions. To fill the critical gap of dual-domain benchmarks, we construct MDHA, a novel dataset for direction-insensitive radar-based HAR. Extensive experiments demonstrate that DIHARNet achieves superior performance, reaching an overall accuracy of 93.46% while maintaining stable recognition across representative aspect-angle variations, highlighting its effectiveness for robust direction insensitive radar-based HAR.