Scholarlink

Wearables equipped with Inertial Measurement Units (IMUs) can be used for Human Activity Recognition (HAR) in real-life scenarios. Typically, activity classification occurs within the microcontroller embedded in the wearable device. In this paper, we demonstrate the feasibility of performing real-time HAR at the extreme edge, hence directly within the smart sensor. The evaluation has been carried out on two 6-axis IMUs developed by STMicroelectronics, namely the LSM6DSV16BX and the LSM6DSO16IS, featuring a Machine Learning Core (MLC) and an Intelligent Sensor Processing Unit (ISPU), respectively. In this study, data were collected by using an earbud, and two tasks were analyzed. The first one focused on dynamic activities (go upstairs, go downstairs, sit on a chair, stand up from a chair, walk, stay still), while the second one led to the recognition of a subject’s posture (sitting still, standing still). The best results for the first task were achieved by the ISPU of the LSM6DSO16IS programmed with a Convolutional Neural Network (CNN) having an accuracy of 99.2% with 3 s long windows as input. The second task was better handled by a decision tree in the MLC of the LSM6DSV16BX, with an accuracy of 79.1% with 4 s long inputs. The two IMUs showed an energy consumption per inference of 930 nWh and 1180 nWh, respectively.

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

Efficient Human Activity Recognition: Machine Learning at the Sensor Level

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Arianna De Vecchi,
  • Alice Scandelli,
  • Federica Bossi,
  • Benedetta Caterina Casadei,
  • Hazem Hesham Yousef Shalby,
  • Marco Boschi,
  • Federica Villa

摘要

Wearables equipped with Inertial Measurement Units (IMUs) can be used for Human Activity Recognition (HAR) in real-life scenarios. Typically, activity classification occurs within the microcontroller embedded in the wearable device. In this paper, we demonstrate the feasibility of performing real-time HAR at the extreme edge, hence directly within the smart sensor. The evaluation has been carried out on two 6-axis IMUs developed by STMicroelectronics, namely the LSM6DSV16BX and the LSM6DSO16IS, featuring a Machine Learning Core (MLC) and an Intelligent Sensor Processing Unit (ISPU), respectively. In this study, data were collected by using an earbud, and two tasks were analyzed. The first one focused on dynamic activities (go upstairs, go downstairs, sit on a chair, stand up from a chair, walk, stay still), while the second one led to the recognition of a subject’s posture (sitting still, standing still). The best results for the first task were achieved by the ISPU of the LSM6DSO16IS programmed with a Convolutional Neural Network (CNN) having an accuracy of 99.2% with 3 s long windows as input. The second task was better handled by a decision tree in the MLC of the LSM6DSV16BX, with an accuracy of 79.1% with 4 s long inputs. The two IMUs showed an energy consumption per inference of 930 nWh and 1180 nWh, respectively.