SKS-transformer: multi-scale and direction-aware attention for inertial sensor-based activity recognition.

Area of Science:

• Computer Science
• Machine Learning
• Wearable Technology

Background:

• Human Activity Recognition (HAR) is crucial for applications like healthcare and sports analytics.
• Scalable and accurate HAR systems that generalize across diverse scenarios remain a challenge.

Purpose of the Study:

• To propose a scalable HAR system integrating a novel SKS-Transformer model with a custom wearable Inertial Measurement Unit (IMU).
• To enhance HAR accuracy and scalability for diverse activities and fine-grained error detection.

Main Methods:

• Developed a custom wearable IMU with ESP8266 and JY61 sensor for wireless motion data acquisition.
• Proposed the SKS-Transformer model featuring Selective Kernel Networks and squeeze-enhanced axial attention for multiscale temporal and directional dependency capture.
• Implemented a data preprocessing pipeline (denoising, segmentation, normalization) and a learnable gating mechanism for adaptive fusion of motion patterns.

Main Results:

• The SKS-Transformer model achieved state-of-the-art performance on public datasets (UCI-HAR, PAMAP2), outperforming 11 other models.
• Attained up to 98.10% accuracy on collected HAR data and 100% accuracy in golf swing error detection.
• Ablation studies validated the contribution of each SKS-Transformer component.

Conclusions:

• The proposed SKS-Transformer model significantly enhances HAR system performance and scalability.
• The system demonstrates high accuracy in recognizing daily activities and detecting specific errors like golf swing faults.
• Future work includes applying the system to intelligent healthcare, sports monitoring, and wearable computing, with publicly available source code.