Dynamic-Attentive Selective Mamba with Group-Aware Convolution for Wearable Sensor-Based Sports and Daily Activity...

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Wearable inertial sensors generate complex motion data crucial for human activity recognition (HAR).
• Current deep learning models for HAR often fail to jointly address body-part specific sensor data, efficient bidirectional temporal modeling, and dynamic attention mechanisms for non-stationary movements.

Purpose of the Study:

• To develop a novel deep learning architecture, Dynamic-Attentive Selective Mamba (DASM), that integrates explicit body-part grouping, linear-time bidirectional temporal modeling, and dynamic attention for HAR.
• To evaluate the performance of DASM against state-of-the-art methods on a benchmark dataset.

Main Methods:

• Proposed DASM architecture combining Group-Aware Convolutions (GroupConv) for body-part feature extraction, Bidirectional Mamba (BiMamba) for efficient temporal context, and Dynamic CBAM (DCBAM) for adaptive attention.
• Experimental validation on the UCI Daily and Sports Activities dataset using stratified segment-level cross-validation and leave-one-subject-out (LOSO) evaluation.

Main Results:

• DASM achieved 99.89% accuracy and F1 score on segment-level validation, surpassing existing models.
• Under LOSO cross-validation, DASM reached 89.34% accuracy, highlighting the challenge of cross-subject generalization.
• Ablation studies confirmed statistically significant performance gains from individual DASM modules.

Conclusions:

• DASM presents a structured deep learning approach that effectively addresses key limitations in current HAR pipelines.
• While achieving near-saturation performance on segment-level tasks, the significant drop in LOSO accuracy indicates a need for further research into robust cross-subject generalization for real-world deployment.