Optimized Support Vector Machine for Human Activity Recognition in Smart Homes

Source abstract

Human activity recognition (HAR) plays a pivotal role in advancing smart home automation, healthcare monitoring, and personalized lifestyle interventions. Traditional HAR systems often struggle with accuracy in diverse environments due to user variability and sensor inconsistencies. This paper introduces an optimized Support Vector Machine (SVM) model enhanced through transfer learning to address these challenges. Leveraging the UCI HAR dataset, the proposed system fine-tunes a base SVM classifier, achieving an overall accuracy of 91.84%, a 10.21% improvement over conventional baselines. Key performance metrics include precision (92.3%), recall (91.8%), and F1-score (92.0%). The optimization incorporates advanced feature extraction, noise filtering, and adaptive fine-tuning, making the system suitable for real-time applications in smart homes. Empirical evaluations demonstrate superior robustness across six activities: walking, walking upstairs, walking downstairs, sitting, standing, and laying. This work contributes to HAR literature by validating transfer learning's efficacy in optimizing SVM for non-invasive, privacy-preserving activity monitoring, with implications for scalable smart environment deployments.