Scholarlink

<p>Motion capture (mocap) systems are extensively utilized in healthcare for monitoring rehabilitation programs, facilitating clinical gait assessments for early Alzheimer’s diagnosis, managing walking disorders, and developing exoskeleton suits. However, like many other healthcare technologies, mocap systems have some flaws, like missing markers and occlusions. Given mocap data’s sequential and temporal nature, understanding marker relationships and capturing global dependencies are crucial for effective human motion recovery applications. To address these challenges, we proposed an unsupervised transformers framework for human motion recovery, called IMU-Trans. We evaluated our framework’s generalizability across two clinical datasets and tested its robustness by adjusting the missing marker rates, comparing its performance against low-dimensional Kalman filtering, long short-term memory (LSTM), and gated recurrent unit (GRU) models. Our experimental results demonstrated that IMU-Trans outperforms state-of-the-art models by training in an unsupervised manner. The closest competitor, GRU, demonstrated an RMSE of 1.35 ± 0.82, 2.36 ± 1.26, 3.43 ± 1.73, and 4.39 ± 2.18&#xa0;cm for 20%, 30%, 40%, and 50% missing rates, respectively. IMU-Trans outperformed GRU with an RMSE of 1.26 ± 0.60, 2.06 ± 0.88, 2.68 ± 1.04, and 3.05 ± 1.22 for the same rates. Notably, our framework performs well even with higher missing data rates, creating opportunities for advancements in data analytics and indicating a promising future for motion capture in healthcare.</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

IMU-Trans: imputing missing motion capture data with unsupervised transformers

  • 期刊论文

  • 发表日期：

    2025年1月3日

查看全文
  • Goksu Avdan,
  • Sinan Onal,
  • Chao Lu

摘要

Motion capture (mocap) systems are extensively utilized in healthcare for monitoring rehabilitation programs, facilitating clinical gait assessments for early Alzheimer’s diagnosis, managing walking disorders, and developing exoskeleton suits. However, like many other healthcare technologies, mocap systems have some flaws, like missing markers and occlusions. Given mocap data’s sequential and temporal nature, understanding marker relationships and capturing global dependencies are crucial for effective human motion recovery applications. To address these challenges, we proposed an unsupervised transformers framework for human motion recovery, called IMU-Trans. We evaluated our framework’s generalizability across two clinical datasets and tested its robustness by adjusting the missing marker rates, comparing its performance against low-dimensional Kalman filtering, long short-term memory (LSTM), and gated recurrent unit (GRU) models. Our experimental results demonstrated that IMU-Trans outperforms state-of-the-art models by training in an unsupervised manner. The closest competitor, GRU, demonstrated an RMSE of 1.35 ± 0.82, 2.36 ± 1.26, 3.43 ± 1.73, and 4.39 ± 2.18 cm for 20%, 30%, 40%, and 50% missing rates, respectively. IMU-Trans outperformed GRU with an RMSE of 1.26 ± 0.60, 2.06 ± 0.88, 2.68 ± 1.04, and 3.05 ± 1.22 for the same rates. Notably, our framework performs well even with higher missing data rates, creating opportunities for advancements in data analytics and indicating a promising future for motion capture in healthcare.