Scholarlink

<p>In recent years, the growing demand for assistive technology in healthcare, fitness tracking and smart environments has positioned Human Activity Recognition (HAR) as vital in the era of data-driven intelligent systems. Identification of motion patterns from various smartphone and wearable sensors is considered to be critical for developing efficient recognition systems. Traditional approaches tend to struggle in capturing complex dependencies in sensor data which leads us to find hybrid techniques for HAR. In this paper, we design a two-stage deep feature extraction and selection technique termed FEST using image encoding derived from time series signal data from sensors. Here, image encoding using Gramian Angular Matrix has been employed to form activity images from sensor readings. These images were then fed to three well-known pretrained visual models, InceptionResNetV2, Xception, EfficientNetV2B1 to capture complex spatial-temporal relations. The deep features were then extracted for optimal feature selection through a two-step selection strategy involving pre-filter Chi-square followed by Honey-Badger meta-heuristic algorithm. The proposed framework has been examined on three public datasets, UCI-HAR, MHEALTH and, KU-HAR which achieved an overall accuracy of 94.00%, 98.12% and 84.04% respectively. The findings underscore the potential of the proposed two-stage visual learning FEST as a reliable solution for sensor-based activity recognition systems.</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

FEST: a deep feature extraction and selection technique for human activity recognition based on smartphone sensor data

  • 期刊论文

  • 发表日期：

    2026年2月10日

查看全文
  • Dipannyta Nandi,
  • Pawan Kumar Singh,
  • Chandreyee Chowdhury

摘要

In recent years, the growing demand for assistive technology in healthcare, fitness tracking and smart environments has positioned Human Activity Recognition (HAR) as vital in the era of data-driven intelligent systems. Identification of motion patterns from various smartphone and wearable sensors is considered to be critical for developing efficient recognition systems. Traditional approaches tend to struggle in capturing complex dependencies in sensor data which leads us to find hybrid techniques for HAR. In this paper, we design a two-stage deep feature extraction and selection technique termed FEST using image encoding derived from time series signal data from sensors. Here, image encoding using Gramian Angular Matrix has been employed to form activity images from sensor readings. These images were then fed to three well-known pretrained visual models, InceptionResNetV2, Xception, EfficientNetV2B1 to capture complex spatial-temporal relations. The deep features were then extracted for optimal feature selection through a two-step selection strategy involving pre-filter Chi-square followed by Honey-Badger meta-heuristic algorithm. The proposed framework has been examined on three public datasets, UCI-HAR, MHEALTH and, KU-HAR which achieved an overall accuracy of 94.00%, 98.12% and 84.04% respectively. The findings underscore the potential of the proposed two-stage visual learning FEST as a reliable solution for sensor-based activity recognition systems.