Cross-Attention Enhanced Pyramid Multi-Scale Networks for Sensor-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for healthcare, smart homes, and gait analysis, with deep learning showing promise.
• A key challenge in HAR is balancing recognition accuracy with computational efficiency, particularly for mobile devices.
• Existing models often struggle to improve feature representation without increasing computational load.

Purpose of the Study:

• To develop a novel deep learning model for Human Activity Recognition (HAR) that optimizes the trade-off between accuracy and computational efficiency.
• To enhance feature representation capabilities without adding significant computational burden for mobile applications.

Main Methods:

• The proposed model integrates a Pyramid Multi-scale Convolutional Network (PMCN) for rich, multiscale feature extraction.
• A Cross-Attention Mechanism is employed to refine interrelationships across sensor, temporal, and channel dimensions, enhancing relevant information.
• The model was evaluated on four diverse datasets: UCI, WISDM, PAMAP2, and OPPORTUNITY.

Main Results:

• The novel HAR model achieved superior activity recognition accuracy across multiple datasets.
• The model demonstrated low computational overhead, making it suitable for resource-constrained environments.
• Ablation and comparative studies confirmed the effectiveness and efficiency of the proposed approach.

Conclusions:

• The developed deep learning model effectively addresses the accuracy-efficiency trade-off in Human Activity Recognition.
• The integration of PMCN and Cross-Attention Mechanism offers a promising direction for efficient and accurate HAR systems.
• This model presents a viable solution for deploying advanced HAR capabilities on mobile and edge devices.