New Algorithm Boosts Athlete Posture Detection Accuracy

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 12 2024

In a paper published in the journal Systems and Soft Computing, researchers proposed a new attitude-solving algorithm to enhance athlete posture detection. They utilized inertial navigation theory, combining complementary filtering and double-layer Kalman filter algorithms to improve accuracy. The model demonstrated high precision in posture detection, offering significant benefits for gymnastics training.

Past work in pose detection technology has largely focused on single techniques like attitude detection or inertial navigation, which have applications in medicine and sports. For instance, many researchers developed a faster OpenPose model for astronaut pose detection, while Yang et al. proposed a fully connected attitude detection network for 3D estimation. However, these studies often needed help with accuracy and stability.

Advanced Athlete Posture Detection

Athletes require precise guidance during training, but traditional instructions are often subjective and lack scientific consistency. A posture detection model has been developed using inertial navigation technology and an advanced attitude settlement algorithm to address this. By determining the navigation coordinate system and the initial alignment of the athlete's posture, the model calculates pitch, roll, and yaw angles using a quaternion method, ensuring accurate and consistent posture tracking.

The detection process starts with the initial calibration of the inertial sensors, where the collected data is processed to calculate the attitude matrix and initialize the quaternion. The angular velocity data from the gyroscope is then adjusted using a complementary filter, combining information from the accelerometer and magnetometer. This step helps to reduce the inherent sensor errors, addressing issues like the gyroscope's zero drift and the accelerometer and magnetometer's limited short-term accuracy.

To further enhance accuracy, the model employs a two-layer Kalman filtering algorithm. The first layer corrects state estimates using roll and pitch angles from the accelerometer, while the second layer refines the yaw angle using magnetometer data. Combining complementary and Kalman filtering algorithms allows for a more precise and stable posture solution, making the model robust in static and dynamic motion scenarios.

Finally, the attitude-solving algorithm is continuously updated, ensuring that the quaternion is always accurate. The entire process, from initial alignment to attitude calculation, is systematically structured to provide real-time, precise posture detection for athletes, supporting their training with scientifically validated guidance.

Athlete Posture Detection

The performance of the athlete posture detection model, based on a posture-solving algorithm, was evaluated through a series of experiments using high-precision gyroscopes, accelerometers, and magnetometers. These sensors were strategically placed on the athlete's back and wrist to capture real-time data during movement.

Data was processed in a MATLAB environment through wireless transmission, with experiments performed in both indoor and outdoor settings to evaluate the model's performance across different conditions. Key parameters, including filter gains and covariance matrices, were meticulously adjusted to enhance the model's accuracy and stability. The study used three filtering algorithms—complementary filtering, an improved two-layer Kalman filtering, and a fusion of both—to process and analyze the data.

Over a 10-minute test period, Algorithm 3 showed a significant decrease in errors, achieving a 35.83% reduction in roll angle error, a 33.67% reduction in pitch angle error, and a 32.73% reduction in yaw angle error compared to the single complementary filtering algorithm. When tested on a turntable at various rotation angles, Algorithm 3 consistently outperformed the other algorithms, demonstrating improved accuracy across all three attitude angles, with an average error reduction of 21.77% compared to the improved Kalman filtering algorithm.

The model's effectiveness was further validated by comparing its performance with other established posture detection models under dynamic conditions. Model 1, based on the study's algorithm, showed the smallest error range in attitude angle detection, maintaining high accuracy and stability throughout the tests.

Model 1 outperformed other models by effectively suppressing gyroscope drift and improving the accuracy of accelerometer and magnetometer data, resulting in superior dynamic performance, faster convergence, and higher resolution accuracy. Additional tests with three athletes confirmed that Model 1 had significantly lower detection errors, making it suitable for real-time posture correction in sports training.

Ablation studies were carried out to evaluate the influence of essential components on the model's performance, including data preprocessing, sensor selection, and filtering techniques. Removing or replacing these components led to a notable increase in detection errors and, in some cases, a slight decrease in solution speed. Notably, the complementary and Kalman filtering algorithms were crucial in minimizing errors and maintaining the model's robustness, as evidenced by the substantial increase in detection errors when these algorithms were omitted.

Conclusion

To sum up, the study constructed a posture detection model for athletes using inertial navigation and posture-solving algorithms. It employed quaternion methods, complementary filtering, and improved Kalman filtering to enhance accuracy, demonstrating a significant reduction in posture angle errors.

Model 1 outperformed others in stability, dynamic performance, and accuracy, making it effective for real-time training assistance. Future research should explore compatibility with various sensors and robustness under diverse environmental conditions.