The Convolutional Neural Networks Training With Channel-Selectivity for Human Activity Recognition Based on Sensors

Area of Science:

• Computer Science
• Machine Learning
• Signal Processing

Background:

• Deep learning models achieve state-of-the-art performance in human activity recognition (HAR) by automatically extracting features from sensor data.
• Training deep neural networks often requires static layers with unchanged weight connectivity, posing computational challenges for mobile platforms with limited resources.

Purpose of the Study:

• To investigate the effectiveness of shallow convolutional neural networks (CNNs) with channel-selectivity for sensor-based HAR tasks.
• To address the computational limitations of deep learning models in resource-constrained environments.

Main Methods:

• Utilized shallow CNNs incorporating channel-selectivity for HAR.
• Conducted extensive experiments on five public benchmark HAR datasets: UCI-HAR, OPPORTUNITY, UniMib-SHAR, WISDM, and PAMAP2.

Main Results:

• Channel-selectivity in CNNs demonstrated lower test errors compared to traditional static layers.
• The proposed channel-selective CNN approach enhances existing deep HAR performance without additional computational cost.

Conclusions:

• Channel-selective CNNs offer an efficient and effective solution for sensor-based HAR on mobile platforms.
• This novel application of channel-selectivity in CNNs advances the field of HAR by improving accuracy and reducing computational demands.