Scholarlink

<p>Human Activity Recognition (HAR) is a promising, rapidly advancing research field that has significant contribution to various real-life applications such as healthcare monitoring, fall detection, sports, and smart home systems. Yet smartphone-based sensors generate high-dimensional data, necessitating effective feature selection methods to maintain recognition performance without the computational burden of redundant information. This paper proposes a novel three-stage filter-wrapper feature selection approach to optimize the recognition of human physical activities. In the first stage, the Pearson Correlation Coefficient (PCC) is calculated to identify and eliminate redundant features, retaining only the most informative ones based on a specified threshold. In the second stage, the remaining features are re-evaluated using the Symmetrical Uncertainty (SU) measure to identify those most relevant to the target classes. The third stage involves refining the SU-selected features through a Genetic Algorithm (GA) wrapper to obtain the final optimal subset. To assess the effectiveness of the proposed method, six classification techniques: Decision Trees (DT), Naive Bayes with Kernel Density Estimation (NB-KDE), Random Forest (RF), Gradient Boosting Trees (GBT), Generalized Linear Models (GLM), and Support Vector Machines (SVM) were employed to discriminate human activities at the first two stages. For the third stage, a SVM-based GA-wrapper was employed to find the optimal feature subset. Extensive experiments were conducted on three benchmark HAR datasets, namely, UCI-HAR, UCI-HAPT, and UCI-AAL, comprising 561 features, and the results demonstrate that the proposed approach significantly reduces the original feature space by 66.31%, 74.15%, and 72.01% while achieving high recognition accuracies of 98.85%, 98.36%, and 93.23% on these datasets respectively. This outcome indicates the superiority of our proposed method over existing approaches.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A three-stage feature selection approach for human activity recognition

  • 期刊论文

  • 发表日期：

    2026年1月2日

查看全文
  • Laith Al-Frady,
  • Ali Al-Taei

摘要

Human Activity Recognition (HAR) is a promising, rapidly advancing research field that has significant contribution to various real-life applications such as healthcare monitoring, fall detection, sports, and smart home systems. Yet smartphone-based sensors generate high-dimensional data, necessitating effective feature selection methods to maintain recognition performance without the computational burden of redundant information. This paper proposes a novel three-stage filter-wrapper feature selection approach to optimize the recognition of human physical activities. In the first stage, the Pearson Correlation Coefficient (PCC) is calculated to identify and eliminate redundant features, retaining only the most informative ones based on a specified threshold. In the second stage, the remaining features are re-evaluated using the Symmetrical Uncertainty (SU) measure to identify those most relevant to the target classes. The third stage involves refining the SU-selected features through a Genetic Algorithm (GA) wrapper to obtain the final optimal subset. To assess the effectiveness of the proposed method, six classification techniques: Decision Trees (DT), Naive Bayes with Kernel Density Estimation (NB-KDE), Random Forest (RF), Gradient Boosting Trees (GBT), Generalized Linear Models (GLM), and Support Vector Machines (SVM) were employed to discriminate human activities at the first two stages. For the third stage, a SVM-based GA-wrapper was employed to find the optimal feature subset. Extensive experiments were conducted on three benchmark HAR datasets, namely, UCI-HAR, UCI-HAPT, and UCI-AAL, comprising 561 features, and the results demonstrate that the proposed approach significantly reduces the original feature space by 66.31%, 74.15%, and 72.01% while achieving high recognition accuracies of 98.85%, 98.36%, and 93.23% on these datasets respectively. This outcome indicates the superiority of our proposed method over existing approaches.