GLULA: Linear attention-based model for efficient human activity recognition from wearable sensors.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Body-worn sensors are crucial for monitoring patient rehabilitation and controlling devices.
• Accurate human activity recognition (HAR) requires capturing complex spatiotemporal data dependencies.
• Optimizing model efficiency (memory, inference time) is vital for real-time and mobile applications.

Purpose of the Study:

• To introduce GLULA, a novel, efficient, and powerful architecture for HAR.
• To address the limitations of existing models in terms of speed and memory usage.
• To enhance HAR performance, especially in data-limited scenarios.

Main Methods:

• Developed GLULA, a unique architecture combining gated convolutional networks, branched convolutions, and linear self-attention.
• Utilized manifold mixup as an augmentation technique to improve performance with limited data.
• Conducted extensive experiments on five benchmark datasets (PAMAP2, SKODA, OPPORTUNITY, DAPHNET, USC-HAD).

Main Results:

• GLULA demonstrated superior performance over recent models on four out of five benchmark datasets.
• The proposed architecture achieved the lowest parameter count among compared state-of-the-art models.
• GLULA exhibited near state-of-the-art inference times, indicating high efficiency.

Conclusions:

• GLULA offers an efficient and effective solution for HAR tasks using body-worn sensor data.
• The architecture's design balances high performance with reduced computational costs.
• GLULA presents a promising advancement for real-time HAR applications in rehabilitation and beyond.