TL;DR
This paper compares four algorithms for calculating running velocity and stride length using inertial measurement units (IMUs) embedded in running shoes. The "Trajectory" algorithm, which reconstructs the foot's path, emerged as the most accurate, performing comparably to GPS for total distance tracking in real-world runs.
What Problem Does This Paper Address?
Tracking running performance, particularly velocity and distance, is crucial for both amateur and professional athletes to monitor progress, prevent overtraining, and enhance training effectiveness. While GPS is a common solution, it has drawbacks like high energy consumption and limited indoor use. Integrating sensors directly into running shoes offers a more convenient and versatile alternative. However, accurately extracting these performance metrics from the complex, dynamic motion of running using only IMUs presents a significant challenge. This paper aims to bridge this gap by evaluating and comparing different algorithmic approaches to determine which best meets the demands of wearable running analysis.
Methods
The study evaluates four distinct algorithms for calculating stride velocity and length from IMU data:
- Stride Time: This algorithm leverages the inverse relationship between stride time and velocity, using stride time and body height to estimate stride length.
- Acceleration: This approach uses foot acceleration data to estimate velocity, correlating an integration value with movement velocity and employing a second-degree polynomial regression.
- Trajectory: This method reconstructs the foot's 3D path using sensor fusion (accelerometer and gyroscope) and linear dedrifting techniques. Stride length and velocity are then derived from this trajectory.
- Deep Learning: A convolutional neural network (CNN) architecture is employed to directly regress stride length from raw IMU data. Velocity is then calculated by dividing the predicted stride length by the stride time.
All algorithms rely on a common stride segmentation method that identifies initial ground contact from acceleration peaks.
Datasets
The study utilized two datasets:
- Lab Study: A controlled environment with 2377 strides from 27 subjects (21 male, 6 female) was used. This dataset was captured using a 3D motion tracking system as the ground truth for velocity and stride length.
- Field Study: A real-world scenario involving 12 subjects performing a self-paced 3.2-km run on a 400-m track. GPS data from a smartphone was used as a reference for total distance.
Key Insights and Ideas
- Trajectory is King (for accuracy): The foot trajectory reconstruction method consistently outperformed other algorithms in the lab study, highlighting the benefit of a more complete understanding of the foot's motion.
- Energy vs. Accuracy Trade-off: The "Acceleration" algorithm offers a compelling balance between accuracy and energy efficiency, requiring only accelerometer data and allowing for lower sampling rates.
- Stride Time's Limitations: The simple stride time-based approach, while energy-efficient, struggles with the complex, non-linear relationships in running biomechanics, leading to lower accuracy.
- Deep Learning's Potential (and Challenges): While deep learning shows promise, its performance in this study was not as robust as the trajectory method, potentially due to variations in running styles and the need for more extensive training data.
- Real-world validation is crucial: The field study confirmed that the Trajectory algorithm's accuracy for total distance tracking is comparable to GPS, demonstrating its viability for practical applications.
Key Results
Lab Study (Mean Error ± Standard Deviation):
- Velocity:
- Stride Time: 0.209 ± 0.782 m/s
- Acceleration: 0.005 ± 0.350 m/s
- Trajectory: 0.028 ± 0.252 m/s (Best)
- Deep Learning: 0.055 ± 0.285 m/s
- Stride Length:
- Stride Time: 17.7 ± 57.3 cm
- Acceleration: -0.5 ± 25.6 cm
- Trajectory: 2.00 ± 14.1 cm (Best)
- Deep Learning: 2.5 ± 20.1 cm
Field Study (Mean Absolute Error for Total Distance):
- Stride Time: 599.7 m
- Acceleration: 333.1 m
- Deep Learning: 194.5 m
- Trajectory: 94.0 m (Best)
- GPS: 82.1 m
Limitations and Future Work
Limitations:
- Trajectory Algorithm's High-Velocity Issue: The Trajectory algorithm's accuracy degrades at higher velocities (>5 m/s) due to the assumption of a zero-velocity phase during midstance, which doesn't hold for fast running.
- Deep Learning Generalization: The DCNN showed varying training and validation errors even after extensive training, suggesting potential generalization issues and the need for more data.
- Stride Time Algorithm's Simplification: The Stride Time algorithm's reliance on a simplified step function and its inability to fully capture velocity control via stride frequency and length limits its accuracy.
- Forefoot Striker Variability: While the Trajectory algorithm performed similarly across strike types, forefoot strikers showed higher standard deviations, likely due to the more dynamic nature of their gait.
Future Work:
- Improving Trajectory Algorithm: Developing better initial condition detection for higher running velocities.
- Enhancing Deep Learning: Incorporating more training data and potentially exploring different architectures to improve generalization.
- Bi-lateral Analysis: Utilizing data from IMUs on both feet to perform more comprehensive analyses, including stride length, time, and orientation imbalances.
- Advanced Temporal Parameters: Investigating the computation of parameters like flight time and stance time, potentially by integrating pressure sensors for more accurate toe-off detection.
Summary updated