Accurate recognition of complex human activities from wearable sensors plays a critical role in sports analytics and human performance monitoring. However, the high dimensionality and redundancy of raw inertial data can hinder model performance and interpretability. This study proposes a hybrid feature selection framework that combines Minimum Redundancy Maximum Relevance (MRMR) and Regularized Neighborhood Component Analysis (RNCA) to improve classification accuracy while reducing computational complexity. Multi-sensor inertial data were collected from field hockey players performing six activity types. Time- and frequency-domain features were extracted from four body-mounted inertial measurement units (IMUs), resulting in 432 initial features. MRMR, combined with Pearson correlation filtering (|ρ| > 0.7), eliminated redundant features, and RNCA further refined the subset by learning supervised feature weights. The final model achieved a test accuracy of 92.82% and F1-score of 86.91% using only 83 features, surpassing the MRMR-only configuration and slightly outperforming the full feature set. This performance was supported by reduced training time, improved confusion matrix profiles, and enhanced class separability in PCA and t-SNE visualizations. These results demonstrate the effectiveness of the proposed two-stage feature selection method in optimizing classification performance while enhancing model efficiency and interpretability for real-time human activity recognition systems.

This study presents a two-stage feature selection framework to enhance the performance and efficiency of human activity recognition (HAR) using inertial sensor data. The proposed pipeline integrates Minimum Redundancy Maximum Relevance (MRMR) with Regularized Neighborhood Component Analysis (RNCA) to systematically reduce feature dimensionality while preserving discriminative power. Initially, MRMR with a correlation threshold (ρ = 0.7) filtered out redundant and irrelevant features, reducing the original 432 features to 222. Subsequently, RNCA refined this subset to 83 features by weighting their joint contribution to classification accuracy. Experimental results, based on a 1-Nearest Neighbor classifier, demonstrated that the RNCA-MRMR model achieved a testing accuracy of 92.92% and an F1-score of 87.13%, slightly outperforming the full-feature model (92.16% accuracy, 86.13% F1-score) while significantly reducing computational complexity. Statistical analysis via paired t-tests confirmed that these performance gains were not statistically significant (p > 0.05), validating that drastic dimensionality reduction did not compromise predictive performance. Visualization through PCA and t-SNE further confirmed improved class separability in the reduced feature space. The resulting model offers enhanced generalization, interpretability, and deployability, making it well-suited for real-time HAR applications on resource-constrained platforms. This study highlights the value of combining filter- and learning-based feature selection to build compact, robust, and high-performing HAR models.