TL;DR
This paper introduces an energy-efficient wearable Electronic Performance Tracking System (EPTS) for football players that uses an on-device Deep Convolutional Neural Network (DCNN) to classify player activities. By dynamically adjusting sensor sampling rates based on these classified activities, the system significantly reduces power consumption without sacrificing data accuracy, extending device battery life by over 1.3 times.
What Problem Does This Paper Address?
Football players increasingly use wearable Electronic Performance Tracking Systems (EPTS) to gather data on their movements and performance. These systems typically rely on Global Navigation Satellite Systems (GNSS) for positioning and Inertial Measurement Units (IMUs) for detailed motion sensing. However, GNSS, in particular, is a major power drain, especially when operating at high sampling rates (e.g., 10 Hz) to maintain accuracy. This high power consumption leads to bulky batteries, limiting player comfort and potentially increasing injury risk. The challenge is to maintain accurate performance data while drastically reducing the EPTS device's energy footprint.
Methods
The researchers developed a prototype EPTS device integrating a GNSS module, an IMU, and a microcontroller. The core innovation lies in an "activity-aware GNSS control" strategy. This strategy employs a compact, on-device DCNN to classify player activities in real-time using IMU data. Based on the classified activity, the system dynamically adjusts the sampling rates of both the GNSS and IMU sensors. For instance, during static or slow movements, the GNSS sampling rate is reduced, and in some cases, IMU operations can be temporarily turned off.
To achieve on-device processing, the DCNN model was heavily optimized:
- Compression: A novel DCNN architecture was designed and then compressed using quantization (8-bit fixed-point) and layer-fusing techniques.
- Custom Dataset: A dataset of football activities was collected using the prototype EPTS devices from five subjects, differing from generic human activity recognition datasets.
- Adaptive Sampling: An iterative algorithm (Algorithm 1) was developed to find near-optimal sensing rate configurations (fG for GNSS, fI for IMU) for each activity, balancing power consumption and measurement error using a cost function.
Datasets
The paper introduces a custom dataset for training and evaluating the DCNN. This dataset consists of 28,712 one-second long samples of IMU measurements from five male subjects performing various football-related activities (walking, jogging, running, stationary, turning). The data was collected using the prototype EPTS devices and includes standardized trajectories and specific tracks designed for linear and turning movements.
Key Insights and Ideas
- Activity-Awareness is Key: The most significant insight is that simply reducing GNSS sampling based on speed (a "straight-forward" method) is insufficient. Accurately classifying the type of activity (e.g., stationary vs. fast running vs. turning) is crucial for intelligent, energy-saving sensor control.
- On-Device DCNN for Efficiency: Deploying a compressed DCNN directly on the EPTS device's microcontroller is feasible and enables real-time activity classification without relying on external processing, which would introduce latency and more power consumption.
- Trade-off Management: The proposed cost function and iterative optimization algorithm effectively balance the competing goals of minimizing power consumption and maintaining acceptable measurement errors (distance and speed).
- IMU Reduction is Also Possible: While GNSS is the primary power drain, the DCNN's ability to classify activities also allows for intelligent reduction of IMU sampling rates, further contributing to energy savings.
Key Results
- Power Savings: The proposed DCNN-based sensing-rate control reduces the overall power consumption of the EPTS device by 28% compared to the baseline operation. This is achieved by reducing GNSS power consumption by 38% and IMU power consumption by 34%.
- Extended Battery Life: The energy savings translate to an extended EPTS device life-time of more than 1.3 times.
- Accuracy Maintained: Despite reduced sampling rates, the system maintains comparable distance and speed measurement errors to the baseline: Ed = 0.071 m and Es = 0.657 m·s⁻¹.
- DCNN Performance: The custom-trained DCNN achieves a high accuracy of 98.29% with a very small model size (7.56 KB for 32-bit floating-point, 1.96 KB for 8-bit fixed-point), enabling on-device deployment.
- Reduced Sampling: The proposed method uses significantly fewer GNSS sampling points compared to baseline and even the straight-forward method, as demonstrated in Figure 8.
Limitations and Future Work
- Dataset Size: While the custom dataset is tailored for football activities, its size and the number of subjects (five) might limit generalizability. Future work could involve expanding the dataset with more subjects and diverse scenarios.
- Real-World Complexity: The paper focuses on standardized testing scenarios and specific football activities. Real match conditions involve more complex and unpredictable movements, which could be a subject for future research.
- Hyperparameter Tuning: The cost function involves hyperparameters (α and β) that balance power and error. While optimized for the turning_slow activity, further investigation into their impact across all activities and potential automatic tuning could be beneficial.
- IMU Interpolation: The paper mentions linear interpolation for IMU data when sampling rates are reduced. Exploring more sophisticated interpolation methods could potentially improve accuracy further.
Summary updated