Scholarlink

The contactless detection of human falls utilizing millimeter-wave radar has gained considerable attention due to its advantages of convenience and high efficiency. This technology holds substantial promise for applications in smart home technologies and intelligent medical systems, and has recently emerged as a key research area. This paper investigates the precise detection of human falls by employing feature spectrograms derived from millimeter-wave radars. Specifically, the IWR1843 radar board captures radar echoes resulting from human falls and conducts a detailed time-frequency analysis of these echoes. Through rigorous data preprocessing, distance-time motion (RTM) and Doppler-time motion (DTM) spectrograms are extracted to furnish a comprehensive representation of human movement characteristics. An innovative integrated feature spectrogram (DTM-RTM) is then formulated to discriminate between six distinct human movements, including three types of falls and three non-fall activities. The performance of three convolutional neural network models—MobileViT, RepLKNet, and TinyViT—was assessed, with each model achieving an outstanding accuracy rate exceeding 98%.

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

Comparing Human Fall Detection Algorithms of Millimeter-Wave Radar Among MobileViT, RepLKNet and TinyViT Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kai Lu,
  • Heng Zheng,
  • Xiangqun Zhang,
  • Dandan Yao,
  • Genyuan Du,
  • Zhizhou Ge,
  • Chenxi Zhao

摘要

The contactless detection of human falls utilizing millimeter-wave radar has gained considerable attention due to its advantages of convenience and high efficiency. This technology holds substantial promise for applications in smart home technologies and intelligent medical systems, and has recently emerged as a key research area. This paper investigates the precise detection of human falls by employing feature spectrograms derived from millimeter-wave radars. Specifically, the IWR1843 radar board captures radar echoes resulting from human falls and conducts a detailed time-frequency analysis of these echoes. Through rigorous data preprocessing, distance-time motion (RTM) and Doppler-time motion (DTM) spectrograms are extracted to furnish a comprehensive representation of human movement characteristics. An innovative integrated feature spectrogram (DTM-RTM) is then formulated to discriminate between six distinct human movements, including three types of falls and three non-fall activities. The performance of three convolutional neural network models—MobileViT, RepLKNet, and TinyViT—was assessed, with each model achieving an outstanding accuracy rate exceeding 98%.