Abstract
The proliferation of collaborative training and multi-person sports has underscored the necessity for concurrent whole-field action sensing. However, Electromyography (EMG) recognition, which plays a pivotal role in Wearable Human Activity Recognition (WHAR) for analyzing muscle activity and decoding action intent, still faces challenges in achieving a balance between performance, cost, and efficiency in multi-person scenarios. Unlike current channel-expansion solutions, we propose a wireless wearable Single-Dimensional Sparse EMG (2SEMG) Sensor for efficient personal sampling. These action-unaffected sensors leverage the proposed lightweight One-Dimensional Motion Network (OMONet) to facilitate concurrent action sensing. Experiments demonstrate that OMONet achieves leading performance and efficiency in action signal recognition, and two real-world badminton matches further confirm the performance, robustness, and real-time efficiency of the whole-field action sensing network constructed via 2SEMG Sensors and OMONet.
Introduction
In collaborative sports like badminton or basketball, tracking every player’s actions simultaneously—referred to as "whole-field sensing"—is technically challenging. Traditional computer vision systems often struggle with visual occlusion, while WiFi-based sensing is prone to signal interference. Electromyography (EMG), which measures muscle electrical activity, offers a direct way to capture action intent. However, existing EMG solutions are often bulky, expensive, and require computationally heavy models that cannot run locally on wearable microcontrollers (MCUs).
System Design
The researchers developed a co-designed hardware-software system to enable real-time, multi-player action recognition using minimal resources.
2SEMG Sensor Hardware
Detailed views of the 2SEMG sensor hardware, showing front, back, and attachment configurations. Source, p. 1
The 2SEMG Sensor is a wireless, single-channel EMG wearable weighing only 7.2g. It operates at ultra-low power (12mW) and supports data transmission up to 100 meters. Its compact design allows it to be worn comfortably on the forearm during intense physical activity without affecting the athlete's range of motion.
OMONet Architecture
The OMONet pipeline processes single-channel EMG signals into action classifications through decoupled motion attention. Source, p. 1
The One-Dimensional Motion Network (OMONet) is a hyper-efficient neural network designed for MCUs. It utilizes a 1D-2D mixed convolution approach to process signals with minimal latency.
- 1D-EMG Perturbation Synthesis (1D-EPS): This module simulates physiological noise during training to improve robustness against real-world signal variations.
- Motion Attention: This core module decouples the EMG signal into Freq-Motion (muscle fiber activation) and Time-Motion (movement continuity) paths, allowing for a significant reduction in model parameters.
- Fusion and Prediction: The decoupled features are compressed and then re-synchronized to produce final action classifications.
Experimental Results
The system was validated using three datasets: the new 1-BEMG badminton benchmark, the Myo gesture dataset, and the Yaseen18 heart sound dataset.
- Accuracy and Efficiency: OMONet achieved a state-of-the-art 99.47% accuracy on the 1-BEMG dataset. Remarkably, it uses only 0.08M parameters, making it approximately 430 times smaller than previous high-performing models like ML-TCN.
- Real-World Performance: Testing during formal badminton matches showed an end-to-end latency of less than 1 second and a recall rate of over 98%.
- Sustainability: The sensor's low power consumption allows for over 10 hours of continuous battery life.
Conclusion
The combination of the 2SEMG sensor and OMONet provides a scalable, efficient solution for whole-field action sensing. By prioritizing hardware-aware model design and feature decoupling, the system achieves high accuracy on low-power edge devices. Future work will focus on expanding the system to track whole-body movements and scaling to larger numbers of concurrent users.
Summary updated