Inertial Sensor-Based Recognition of Field Hockey Activities Using a Hybrid Feature Selection Framework.

Area of Science:

• Sports Science and Biomechanics
• Machine Learning and Data Science
• Wearable Technology and Sensor Systems

Background:

• Accurate human activity recognition using wearable sensors is crucial for sports analytics and performance monitoring.
• High dimensionality and redundancy in raw inertial data pose challenges for model performance and interpretability.
• Existing methods often struggle to balance classification accuracy with computational efficiency.

Purpose of the Study:

• To propose and evaluate a hybrid feature selection framework combining Minimum Redundancy Maximum Relevance (MRMR) and Regularized Neighborhood Component Analysis (RNCA).
• To enhance classification accuracy and reduce computational complexity in human activity recognition from multi-sensor inertial data.
• To improve the interpretability and efficiency of models for real-time applications.

Main Methods:

• Collected multi-sensor inertial data from field hockey players across six activity types.
• Extracted time- and frequency-domain features from four body-mounted inertial measurement units (IMUs), yielding 432 initial features.
• Implemented a two-stage feature selection: MRMR with Pearson correlation filtering, followed by RNCA for supervised feature weighting.

Main Results:

• The hybrid framework achieved a test accuracy of 92.82% and an F1-score of 86.91% using only 83 selected features.
• The final model outperformed configurations using MRMR alone and slightly surpassed the performance of the full feature set.
• Demonstrated reduced training time, improved confusion matrix profiles, and enhanced class separability via PCA and t-SNE visualizations.

Conclusions:

• The proposed two-stage feature selection method effectively optimizes classification performance for human activity recognition.
• This framework enhances model efficiency and interpretability, making it suitable for real-time systems.
• The hybrid approach successfully addresses the challenges of high dimensionality and redundancy in wearable sensor data.