TL;DR
This paper introduces a novel wearable sensor system combined with deep learning for real-time biomechanical feedback in hammer throw training. By miniaturizing hardware and leveraging AI to estimate complex joint angles from limited sensor data, the system offers a practical, non-intrusive solution to improve athlete performance and training efficiency.
What Problem Does This Paper Address?
Traditional biomechanical analysis for sports like hammer throw often relies on expensive, lab-based optoelectronic systems that are impractical for real-world training environments. While wearable sensors offer a more practical alternative, existing systems often require a large number of sensors, leading to movement constraints and reduced athlete comfort. Furthermore, accurately quantifying complex, high-speed movements like the hammer throw using limited wearable sensors has been a significant challenge. This paper aims to bridge this gap by developing a compact, user-friendly wearable system that uses AI to infer crucial biomechanical parameters from minimal sensor input, enabling real-time feedback for athletes.
Methods
The core of this work lies in the integration of a custom-designed wearable sensor system with deep learning algorithms.
- Wearable Sensor System: The system comprises a miniaturized sensor node featuring an Arduino microcontroller, an XBee wireless communication module, two Inertial Measurement Units (IMUs), and a micro load cell. The IMUs are strategically placed to capture wrist and hip movements, while the load cell measures wire tension in the hammer. This design is a significant improvement over previous iterations, reducing the device's size and complexity.
- Deep Learning Models: Two deep neural network (DNN) models were developed using TensorFlow and Keras.
- Simplified Model: Estimates selected lower limb joint angles (hip, knee, ankle) using wrist and hip position/velocity data.
- Complete Model: Estimates a broader range of joint angles, including upper and lower limbs, using the same input data.
- Both models utilize ReLU activation functions and the Adam optimizer, trained for 1000 epochs. Madgwick's filter is used for IMU data fusion.
Datasets
The datasets used for training and evaluation were collected from a single college-level athlete performing hammer throws. Motion capture data from a Vicon® 3D motion capture system served as the ground truth for kinematic data. The paper does not introduce a new benchmark dataset but rather utilizes existing motion capture technology for data collection.
Key Insights and Ideas
- AI as a Sensor Multiplier: The study demonstrates that AI, specifically deep learning, can effectively "multiply" the information gained from a limited number of wearable sensors. By learning complex relationships between sensor data and biomechanical parameters, AI can infer information that would typically require many more sensors.
- Practicality Over Sensor Count: The research prioritizes practicality by minimizing the number of wearable sensors. The system's design ensures minimal interference with the athlete's natural movement, a critical factor for effective training.
- Synchronized Kinematic and Kinetic Data: For the first time, the system allows for the synchronous collection of both kinematic (joint angles from IMUs) and kinetic (wire tension from load cell) data. This integrated approach is crucial for a comprehensive understanding of motor control and for developing personalized training strategies.
- Two Distinct Tension Patterns: The analysis of the load cell data revealed two distinct patterns of tension development during the throw, suggesting different underlying biomechanical strategies that can be further investigated with AI.
Key Results
- Load Cell Accuracy: The load cell achieved a mean relative error of 0.87% after calibration, indicating high linearity and accuracy for measuring wire tension.
- IMU Accuracy: The IMUs demonstrated an accuracy of 6% after calibration against a Vicon® system, deemed sufficient for sports skill analysis.
- DNN Model Performance:
- The complete DNN model significantly outperformed the simplified model in estimating joint angles.
- For the complete model, the Mean Absolute Errors (MAEs) were approximately ±4° for hip, ±4° for knee, and ±5° for ankle flexion/extension angles.
- The simplified model had higher MAEs, around ±12° for hip, ±14° for knee, and ±11° for ankle.
- Data Normalization Impact: Normalizing the datasets for the DNN models, while improving accuracy metrics, paradoxically reduced the diversity of the data, potentially hindering the models' ability to fully capture complex movements.
Limitations and Future Work
- Data Volume: A significant limitation is the limited volume of training data, collected from a single athlete. This restricts the DNN models from fully revealing their potential and may contribute to outliers.
- Single Athlete Data: The reliance on data from only one athlete raises concerns about generalizability. Future work should involve collecting data from a larger and more diverse group of athletes.
- Outlier Management: The presence of outliers in the simplified model's training data negatively impacted its performance. Developing algorithms to effectively remove or mitigate these outliers is a key area for improvement.
- Further Sensor Integration: The authors suggest exploring the integration of additional sensors, such as an upper-body IMU or surface electromyography (sEMG) sensors, to capture a more comprehensive biomechanical picture.
- Vertical GRF Measurement: Integrating a vertical Ground Reaction Force (GRF) measurement module is planned for future iterations.
- Soft and Flexible Sensors: Investigating soft and flexible IMU sensors is proposed as a way to further enhance practicality and reduce movement constraints.
- Generalizability to Other Sports: The framework developed in this study is intended to be adaptable for developing wearable biomechanical feedback systems for other elite sports.
Summary updated