Wearable inertial sensors produce multi-axis motion signals with rich spatial and temporal structure. Existing deep-learning pipelines for human activity recognition (HAR) rarely tackle three issues jointly: explicit modeling of the body-part grouping of multi-location inertial channels, bidirectional temporal modeling at linear-time cost, and dynamic, time-varying attention for non-stationary motion. We aim to close these three gaps within a single architecture. To this end we propose Dynamic-Attentive Selective Mamba (DASM), which combines three components: Group-Aware Convolutions (GroupConv) for body-part-aware local features, a Bidirectional Mamba (BiMamba) module for linear-time forward and backward temporal context, and a Dynamic CBAM (DCBAM) that produces per-timestep channel and spatial attention for non-stationary windows. On the UCI Daily and Sports Activities dataset (19 classes, 8 subjects), under stratified segment-level 5-fold cross-validation (3 seeds, 15 runs/model), DASM reaches 99.89% accuracy and F1, a 0.11% gain over CNN-BiGRU-CBAM and 0.50% over Multi-STMT; under leave-one-subject-out (LOSO), it reaches 89.34%, 1.69% above the strongest baseline. The 10.55% drop under LOSO shows that segment-level results overestimate cross-subject generalization. Ablations show small but statistically detectable gains (Cohen’s d∈0.4,0.7 per module, d≥1.5 full-vs-baseline). We therefore position the contribution as a structured architecture within a near-saturated benchmark; broader deployment claims require multi-dataset subject-independent validation.
Z et al. (Sat,) studied this question.