Scholarlink

<p>Human locomotion recognition (HLR) is essential for the self-paced treadmill and other human-robot interactive systems. The generalization capability of recognition algorithms should be carefully considered due to the diversity of locomotion patterns for different subjects. Conventional CNN- or LSTM-based pipelines suffer a marked loss of accuracy once a new treadmill user is encountered, and domain-adaptation approaches such as DANN still rely on collecting unlabeled data from that user. In response, we previously introduced a Hybrid Spatial–Temporal Graph Convolutional Network (HSTGCN) that preserves the natural topology of plantar-pressure sensors through an adaptive spatial graph, extracts modality-intrinsic features by processing pressure data with a spatial–temporal GCN and inertial data with an LSTM, and then fuses these heterogeneous streams through a temporal LSTM to produce a compact, user-invariant representation of locomotion. The present study is the first to rigorously assess whether these architectural choices truly deliver ready-to-use generalization. The proposed HSTGCN is validated on a dataset consisting of eight subjects with five locomotion modes. Under this test, the HSTGCN retained 97.9%±1.4% accuracy−7.6 percentage points above the CNN and statistically indistinguishable from a DANN baseline, yet without requiring any target-subject data. Confusion-matrix inspection confirmed per-mode recall above 95%, while t-SNE visualizations revealed that only the HSTGCN produced clusters that were well separated by class yet overlapped across subjects, explaining its user-independent behaviour. Ablating either the modality decomposition/late fusion module or the graph-based spatial extractor reduced accuracy by up to 4% and tripled inter-subject variance, pinpointing the mechanisms that underwrite the model’s robustness. Together these findings demonstrate that HSTGCN offers a user-independent, ready-to-use solution for next-generation self-paced treadmills as well as for other wearable-sensor locomotion 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

Multimodal Sensor Insoles Based User-Independent Human Locomotion Recognition for the Self-Paced Treadmill

  • 期刊论文

  • 发表日期：

    2026年4月23日

查看全文
  • Sen Lu,
  • Yu-Yang Qian,
  • Kai-Ming Yang,
  • Yu Zhu

摘要

Human locomotion recognition (HLR) is essential for the self-paced treadmill and other human-robot interactive systems. The generalization capability of recognition algorithms should be carefully considered due to the diversity of locomotion patterns for different subjects. Conventional CNN- or LSTM-based pipelines suffer a marked loss of accuracy once a new treadmill user is encountered, and domain-adaptation approaches such as DANN still rely on collecting unlabeled data from that user. In response, we previously introduced a Hybrid Spatial–Temporal Graph Convolutional Network (HSTGCN) that preserves the natural topology of plantar-pressure sensors through an adaptive spatial graph, extracts modality-intrinsic features by processing pressure data with a spatial–temporal GCN and inertial data with an LSTM, and then fuses these heterogeneous streams through a temporal LSTM to produce a compact, user-invariant representation of locomotion. The present study is the first to rigorously assess whether these architectural choices truly deliver ready-to-use generalization. The proposed HSTGCN is validated on a dataset consisting of eight subjects with five locomotion modes. Under this test, the HSTGCN retained 97.9%±1.4% accuracy−7.6 percentage points above the CNN and statistically indistinguishable from a DANN baseline, yet without requiring any target-subject data. Confusion-matrix inspection confirmed per-mode recall above 95%, while t-SNE visualizations revealed that only the HSTGCN produced clusters that were well separated by class yet overlapped across subjects, explaining its user-independent behaviour. Ablating either the modality decomposition/late fusion module or the graph-based spatial extractor reduced accuracy by up to 4% and tripled inter-subject variance, pinpointing the mechanisms that underwrite the model’s robustness. Together these findings demonstrate that HSTGCN offers a user-independent, ready-to-use solution for next-generation self-paced treadmills as well as for other wearable-sensor locomotion systems.