Deep learning algorithms have demonstrated remarkable efficacy in HAR systems, delivering superior recognition results compared to traditional methods. However, their reliance on extensive data and high computational power poses significant challenges, particularly for deployment on resource-constrained edge devices. In contrast, shallow and ensemble learning algorithms offer a more practical alternative, requiring less computational overhead while still achieving competitive performance. This chapter introduces a novel data intensity-based feature selection protocol designed to optimize sensor-based HAR systems by systematically identifying and retaining the most impactful features while discarding redundant or less informative ones. The proposed method leverages Root Mean Square (RMS) to evaluate the statistical significance and discriminative power of features based on their intensity across various activities of daily living. By focusing on feature dependencies in 2D and 3D sensor data, the protocol ensures the removal of low-intensity features without compromising the relationships between variables, thereby reducing dimensionality and mitigating overfitting. Experimental validation on three datasets—our custom dataset, MotionSense, and mHealth—demonstrates that the proposed method achieves comparable accuracy, precision, recall, and F1-score while significantly optimizing computational efficiency. The results confirm that the removal of low-intensity features reduces computational time by up to 65.8% for shallow learners and 48.2% for ensemble models, without sacrificing performance. This chapter highlights the potential of data intensity-based feature selection to enhance the scalability and efficiency of HAR systems, particularly for real-time applications on wearable and edge devices. Future work will explore automated feature intensity analysis to further streamline the feature selection process.