Scholarlink

Abstract <p>The problem of increasing the accuracy of determining the position of a patient’s head during the vHIT medical test using video recordings obtained from a smartphone camera is considered. The focus is on comparing the output of three machine-learning models (Dense Head Pose Estimation, Head Pose Estimation, and Lightweight Head Pose Estimation) designed for head pose detection with reference data obtained using IMU sensors. To improve the accuracy of signals obtained from a video stream, four digital filtering methods were compared, namely, empirical mode decomposition (EMD), variational mode decomposition (VMD), discrete wavelet transform, and Butterworth filtering. The mean square error (MSE), weighted by the signal amplitude, was used as a quality metric. The results showed that the best fit to the benchmark data was achieved by using the LHPE model in combination with the wavelet transform and Butterworth filtering. The presented approach confirms the feasibility of using a smartphone camera for high-precision head detection, which opens up the prospect of using ordinary smartphones as hardware for the vHIT test.</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 Comparison of Data-Filtering Methods Obtained Using Different Machine-Learning Models for Head Detection during vHIT Recordings on a Smartphone Camera

  • 期刊论文

  • 发表日期：

    2026年2月1日

查看全文
  • G. D. Mamykin,
  • A. A. Kulesh,
  • F. L. Barkov,
  • V. P. Pervadchuk,
  • A. O. Oskolkova,
  • S. Krapivin,
  • Yu. A. Konstantinov

摘要

Abstract

The problem of increasing the accuracy of determining the position of a patient’s head during the vHIT medical test using video recordings obtained from a smartphone camera is considered. The focus is on comparing the output of three machine-learning models (Dense Head Pose Estimation, Head Pose Estimation, and Lightweight Head Pose Estimation) designed for head pose detection with reference data obtained using IMU sensors. To improve the accuracy of signals obtained from a video stream, four digital filtering methods were compared, namely, empirical mode decomposition (EMD), variational mode decomposition (VMD), discrete wavelet transform, and Butterworth filtering. The mean square error (MSE), weighted by the signal amplitude, was used as a quality metric. The results showed that the best fit to the benchmark data was achieved by using the LHPE model in combination with the wavelet transform and Butterworth filtering. The presented approach confirms the feasibility of using a smartphone camera for high-precision head detection, which opens up the prospect of using ordinary smartphones as hardware for the vHIT test.